Agentic AI Platform Change Log

MCP / tool calling / API

Smithery's connection management services shut down on October 31, 2026. Server URLs on run.tools and server.smithery.ai, the Smithery API, Connect and Toolbox, Smithery accounts and API keys, and the CLI and SDKs all stop working, with current CLI and SDK versions marked deprecated. What remains is a public directory of MCP servers, powered by Arcade.dev, that no longer proxies connections and lists each verified server's direct MCP URL.

Agents that reach tools through Smithery URLs or its API need each vendor's official MCP URL, shown on the server's Smithery page, and products that manage MCP connections for their own users need a new connection manager. Paid plan charges since August 5 are refunded and subscriptions canceled.

Security / enterprise

Claude Code 2.1.316 adds pooled group budgets to the Claude apps gateway, so a group's members share one monthly budget on top of each person's own limit. A new layered policy mode applies every gateway policy that matches a developer, merged in order, instead of only the first match. Bedrock upstreams can also tag prompts so an Amazon Bedrock guardrail's prompt attack filter can screen Claude Code traffic.

Platform teams rolling Claude Code out across many developers can cap spend by team and stack company wide rules with team rules, without building their own controls.

Deployment / data residency

The new Dock CLI in Capy 0.7.0 runs Capy without the desktop app, so a build server, cloud VM or spare Mac can host agent threads with the tools and credentials already on it. The same release moves Captains and Issues out of experiments and turns them on for every organization. A Captain directs a crew of threads, and Issues tracks the work they pick up, with blocking links, sub issues, labels, due dates and linked pull requests.

Engineering teams can run agent work on their own machines with no screen attached, and get a shared tracker that agents update as they go, which makes a crew of agents easier to supervise.

Integrations

Basis in your binder runs the Basis audit agents inside the engagement binder a firm already uses, where before they ran in a separate window beside it. Auditors start at review, with no documents to move between tools and firm standards applied across every engagement.

Audit practices can put agents on the first pass of busy season work without a binder migration or a change in where staff work, which removes a common reason to delay adoption.

Browser/computer use

BU Ultrafast is open to every Browser Use Cloud plan in the US, including the free plan, and its waitlist is closed. Developers switch to it by setting the model to bu-ultrafast on a run, with no other model settings to tune. A companion option runs the same agent with more reasoning per step.

US teams can try the Ultrafast agent on their own browser tasks without waiting for access, and choose per run between speed and extra reasoning.

Memory / state

A new Sessions, Memory and Usage tab in CodeGPT 3.24.80 lists the sessions saved by every coding agent on the computer, including CodeGPT, Claude Code, Codex and Copilot. The list is searchable, shows the memory files behind those sessions alongside the organization's usage, and opens from a new bar at the top of the chat.

Developers who switch between several coding agents can find past work and check what each agent remembers in one place, with the organization's usage in the same view.

Agent capability

The Request Analysis Agent now marks request documents that only partly satisfy the request details in yellow. Its AI summary explains which information is missing, so reviewers can see what still needs to be provided.

Audit teams get a middle state between satisfied and not satisfied, which makes it faster to spot the evidence that needs follow up.

Pricing / packaging

Review usage in ThinkReview version 2.3.14 now runs on a monthly credit pool plus a daily burst cap. When a user hits a limit, the upgrade screen says which one, how many credits have been used and when they reset, with options that match the plan: Free, Lite, Professional or Teams. Monthly and daily usage meters also sit in the review panel's settings menu.

Teams deciding how much code review to run through ThinkReview can see both the monthly and daily limits on their plan before reviews stop, rather than finding out when a review is blocked.

MCP / tool calling / API

Version 0.119.0 of the Zapier SDK adds runWorkflow, a beta method that fires a published workflow's trigger and returns the finished run with its output, or with its error if the run failed. Other workflow run methods can now also wait for a run to end instead of returning at once.

Developers and coding agents can call a Zapier workflow and get its result back in one step instead of polling for it. The method is in beta, so changes can still arrive in minor releases.

Security / enterprise

Around the clock threat hunting from CrowdStrike's OverWatch team now covers AI applications and the autonomous agent activity that Falcon Guardian monitors. Falcon Complete for Falcon Guardian adds 24/7 managed detection and response for AI coding assistants and agents, with analysts who investigate malicious prompts and suspicious agent behavior and then contain confirmed attacks. The same update brings OverWatch hunting on Google Cloud telemetry up to the depth it already had on AWS and Azure.

Security teams that let coding agents loose on endpoints can hand watching and responding to that activity to CrowdStrike's own analysts instead of staffing it themselves.

Agent capability

Lumi, the Luminance assistant, now writes comments inside Word that explain to the counterparty the commercial reasoning behind each proposed redline. Lawyers talk the business's negotiating position through with Lumi, and it drafts the explanation into the document.

Negotiators spend less time justifying each change by hand, and the other side sees why a clause matters, which can cut a round of back and forth.

Agent capability

When Claude or Codex needs a decision in the middle of a run, it now asks inside the Tembo session, and the user picks one of its options or types an answer so the agent can carry on. Projects can also keep a few sandboxes warm on a schedule, such as weekdays 9 to 5, so new sessions start instantly. That warm pool compute appears on the usage page.

A person can steer a coding agent at the moment it is unsure instead of restarting the run, and teams that work set hours can trade a visible compute cost for no wait at session start.

Integrations

A Google Drive connector now works with Figma Make and the Figma agent, so users can summarize research, reference real data and pull in assets from Drive, or create new Drive files, without leaving their file. Connectors are now open on every Figma plan, including the free Starter plan, and are added from the plus menu in the prompt box.

Teams can ground Make prototypes in the documents and data they already keep in Drive, and the move to all plans lets anyone try connectors without an upgrade.

Browser/computer use

Scenarios in Make can now include Claude browsing the web, through a new Browser Use Tool option in the Anthropic Claude app's Create a Prompt module. The model can navigate websites, extract text and manage browser tabs as a step in an automation. New fields carry the browsing session from one scenario step to the next, and a tool result can now be flagged as an error so the model gets better feedback.

Automations can now reach websites that have no API or Make connector, and keep one browsing session across several steps for interactive web tasks.

Security / enterprise

New tables in Bolt database are now private by default, so an app needs explicit permission to read or write their data. The change covers tables created from October 30 and leaves existing databases as they were. Bolt sets up the permissions when it creates a database, and users can ask it to fix a table if a permission error appears.

Apps built with Bolt no longer leave fresh tables open by default, which closes a common route to accidental data exposure in apps written by an agent.

Human approval / guardrails

Governed mode replaces the single Endorsed setting with rules admins write themselves about which warehouse tables, semantic models and projects the Hex Agent may use, for Explorers or for the whole workspace. An admin can limit the agent to semantic models only, to projects only, or to a mix such as all semantic models plus Gold tables and Certified projects. Existing Endorsed mode settings carry over unchanged.

Data teams can keep the agent answering from trusted metrics and certified work instead of raw tables. It governs what the agent uses, not what people can reach, so warehouse permissions still do that job.

Agent capability

Release 6.2.15 of the Retell SDK adds duplex, a new response engine for Retell LLM agents that pairs the gpt-live-1 realtime speech model with a text LLM backend, an optional backup and a voice prompt of up to 15,000 tokens. In duplex mode, tools always speak after they run. The same release adds an API call to refresh a knowledge base and a way to list the catalog of integration app templates.

Voice agent teams can try live speech to speech conversation while keeping a text model behind it for tool use, and can keep knowledge bases current through the API instead of rebuilding them.

Integrations

A new Salesforce SMTP Connector hands email conversations from an Ada AI Agent to Salesforce as Cases, with the transcript and the customer's latest attachments. Human agents reply from the Case using the same support address, so the customer stays in one email thread, and those replies show up in Ada transcripts and reports. One Salesforce org can take handoffs from more than one AI Agent in the same region.

Service teams on Salesforce can escalate email from the AI Agent into their normal Case workflow without a second reply address or a gap in reporting.

Agent capability

Drafts and version history are now on all Chatbase accounts. Saved edits go into a draft while customers keep talking to the live agent, and each publish is stored as a version with highlighted diffs and restore at any time. Backstage, the built in operations assistant, can also work on a draft and restore earlier versions with the user's approval.

Teams can change an agent's instructions, model or actions and test them in the Playground or Tests before customers see anything, then roll back fast if a change goes wrong.

Deployment / data residency

Brands with more than one Shopify store can now run them all from a single Consio workspace. Campaigns, phone numbers, AI agents and ecommerce integrations each belong to a store, while knowledge sources can serve one store or be shared across all of them. A new AI agents page shows each agent's type, store and call sources, and every action opens a panel that explains what it does and the guardrails it follows.

Merchants with several brands can give each store its own voice agents and numbers without juggling separate workspaces, and can check what each agent is allowed to do on a call.

Agent capability

FlowX.AI 6.0 upgrades the Developer Agentic Harness, the Designer's AI agent that plans app changes and builds them on its own branch. A plan can now be split into stories that the agent builds one slice at a time, each with steps to test it before the next one starts. Comment mode lets a builder select a component in the running app and say what should change, and a single agent rail replaces the Command Center and the AI buttons in each editor.

Builders check small working pieces instead of one large agent run, and the agent asks before it reads beyond its target project or serves an app. The release is live on FlowX.AI SaaS, while self hosted customers usually get it with the next LTS release.

Observability / auditability

A new AWS CloudWatch agent for Harness runtime protection, first released as version 0.1.0, discovers AI agents running on AWS Bedrock and collects data from them, including CloudWatch log events. It covers agents built on Bedrock AgentCore as well as standard Bedrock agents. The Terraform and CloudFormation templates for it no longer install the bundled platform agent by default.

Security teams can bring AI agents running on Bedrock under the same Harness runtime protection as their APIs, without writing their own log collection.

MCP / tool calling / API

In public beta, Super Admins can now let supported AI connectors read Sensitive Data in HubSpot accounts that have Sensitive Data turned on, which connectors could not reach before. ChatGPT, Claude and Copilot can read Sensitive Data properties and engagements, while Gemini and the Remote MCP server can read engagements only. Access is read only, users still need their own CRM permissions to see the data, and Highly Sensitive Data properties stay off limits.

Teams in privacy heavy accounts can use outside AI assistants on records that were blocked until now, with admins deciding which connectors get access and each user's permissions still applying.

Agent capability

Enterprise teams can now describe a change and have a Replit Cloud Agent find the relevant code, make the change and test it in the running app. The agent then opens a GitHub pull request or GitLab merge request with screenshots and video. Work runs in isolated cloud sandboxes built from a shared development environment the team sets up once, and follows the team's contribution guidelines.

People who are not engineers can propose tested changes to the company's real codebase, while engineers still review every change before it ships.

Browser/computer use

When asked in web chat or Slack, an Artisan agent in Artie now creates website accounts with its own email and profile instead of a person's logins. Account creation works with signups that send an emailed code and with sites that ask for a password before verification.

An agent that signs up for web tools under its own identity keeps its activity apart from employees' personal accounts, so nobody has to hand over a login for it to work on a new site.

Browser/computer use

Portal Agent v2 works from a Chrome side panel, reading a questionnaire portal page by page and choosing its own next step as it imports the whole DDQ into AutoRFP.ai with question numbers, sections, multiple choice and follow up questions intact. After the user reviews the import, it writes approved, source cited answers back into the portal fields or out to a Word template, and it never submits the questionnaire itself. Imports pick up where they left off after a refresh or a closed tab, and the agent is tested on Dasseti, DiligenceVault, CyberGRX, OneTrust and UpGuard.

Teams answering investor and security questionnaires in portals can drop the export and copy back steps where follow up questions go missing, while a person still reviews the import and decides when to submit.

Agent capability

Managed Agents, Cartesia's hosted voice agent platform, is now available, with its voice models and agent harness designed together so listening, turn taking, tool calls and speech work as one system. It bundles telephony, knowledge bases, backend tool calls and outbound batch campaigns, and agents can switch between English, Spanish, Hindi, French and Japanese mid conversation. Launch pricing is $0.03 a minute, which is 50% off.

Buyers get telephony, knowledge bases, tool calls and outbound campaigns on one platform rather than a voice pipeline they assemble from separate vendors.

MCP / tool calling / API

Crush v0.99.0 widens hooks so teams can run their own checks on session, turn and permission events, and validates hook event names against a known list. The release adds Catwalk v3 support, where custom model providers are defined by the API format they speak rather than by vendor name, and existing configs migrate on their own. An app that embeds Crush can also answer the agent's permission prompts itself.

Teams can attach their own policy checks and logging to the moments when the agent starts work or asks for approval, and editors or internal tools that embed Crush can keep those approvals in their own interface.

Agent capability

Sessions in Jules can continue until November 11, 2026, after which Jules turns read only, and the product winds down for good on December 31, 2026 as Google unifies its agentic coding work. Users can now export their settings, environment variables, scheduled tasks, connected repositories, memories, setup scripts and full session history, including plans, activity logs and code patches, to a local folder through the Jules Tools CLI. Code, branches and pull requests already live in GitHub and stay there.

Teams using Jules for asynchronous coding or scheduled maintenance have until November 11 to finish work in progress and need another coding agent before year end. The export keeps their task history and setup once Jules is gone.

Browser/computer use

Enterprise Browser for Agents lets endpoint and cloud agents work in SaaS, internal and public web apps through Island's browser, with the organization's policies applied. Admins set per agent and per application which pages and DOM operations an agent may use and whether it may write or only read, and the MCP layer follows the same policy. Audit logs record whether a human or an agent acted, sessions can be recorded for playback, and a built in vault fills in logins so passwords never reach the model's context window.

Security teams can let agents into business apps while keeping systems of record read only and keeping an audit trail that separates human actions from agent actions.

Browser/computer use

Locally run browser sessions in Notte SDK v1.10.2 can now use the act, choose and check steps. When an act step's own action opens a JavaScript dialog, the agent can answer it, and it decides again if a dialog covered its target while it was deciding. Clicks also follow the tab they leave the browser on and wait for a page that reloads.

Teams running web agents on their own machines can use these action steps there, and fewer runs stall on pop up dialogs, new tabs and page reloads.

Workflow orchestration

Swarms in Grep AI, formerly Parcha, are teams of agents in which a coordinating agent breaks a task into parts, hands them to specialist agents with their own instructions, skills and tools, and combines their findings for review. Uses include merchant categorization, account research, vendor review and customer onboarding with KYB and KYC tasks, run across a list. Agent code runs in isolated environments, and customers control tool and credential access and can set cost limits for each run.

Compliance and operations teams can hand a repetitive backlog to coordinated agents with scoped access, visible activity and spend caps, while decisions that need judgment stay with people.

MCP / tool calling / API

A new ultra-preview effort level in the Perplexity Agent API handles broad research and data gathering jobs where coverage matters most, and its runs can take a long time. The xhigh and high levels now allow 300 and 40 steps per run, up from 100 and 15, and callers can set their own step limit as high as 500. The undocumented ultra name has been removed and now returns an error.

Teams building research agents on the API can let one run take far more steps before it stops, which suits jobs that pull data from many sources. Anyone still calling the old ultra name must switch.

MCP / tool calling / API

Upstream MCP servers that use the OAuth client credentials grant can now sit behind Portkey's MCP Gateway, starting with Enterprise Gateway 2.32.0. Admins enter the client ID, client secret, token endpoint and an optional scope, and the gateway requests only that scope and fetches and refreshes tokens on its own. The same release lets the Agent Gateway pass chosen headers from the caller's request through to the upstream agent.

Platform teams can route agents to machine to machine MCP servers through the gateway without writing and maintaining their own token handling.

Integrations

Techtouch AI Hub now connects to Jichitai AI zevo, the generative AI platform Shiftplus offers to local governments, so staff can use it from inside the screens of their business systems. The connection works in LGWAN environments and has been available since October 26, 2026. It was built jointly with Miyakonojo City in Miyazaki Prefecture and Shiftplus.

Local governments that already use zevo can put it in front of staff inside the business systems they work in every day, rather than in a separate AI tool.

Integrations

A new 1Password integration lets a Townie sign in to websites with the user's 1Password logins, so tasks that sit behind a login page get finished. The user approves the logins each task needs inside 1Password, and Town cannot see the credentials.

The assistant can now finish work on sites that need a sign in without holding passwords itself, with the user approving each login. Only Personal vault logins on US region Individual and Families accounts work for now.

Pricing / packaging

Free Mode and the Free Trial are retired in Verdent Desktop v2.18.0, along with the 100 welcome credits. GLM 5.3 Flash and DeepSeek V4 Flash stay available but are no longer free, and continued use of the AI features now runs through a paid subscription with Eco Mode and credits.

Teams that were trying Verdent on the free tier now need a paid plan to keep using its AI features, so trial and budget plans need a fresh look.

Agent capability

Better Stack's AI SRE now runs inside Claude as a plugin, free on every Better Stack plan and installed from Claude or Claude Code. A Claude routine fired by a Better Stack incident webhook can go on call and return a root cause analysis, a mitigation plan and a pull request with a fix. A scheduled Claude task can also check errors every few hours and fix them through pull requests, and Claude can set up monitors, dashboards, alerts, heartbeats and status updates on request.

Teams already paying for Claude get an on call agent that works from their own telemetry and code at no extra cost, while Codex and Cursor users still have to wait.

Integrations

The Text AI Agent can now recommend products and show product cards in chat straight from a store's Google Shopping feed. Merchants paste their Google Merchant Center feed link into the Recommend products skill, and the agent suggests only items that are in stock. The feed refreshes every hour.

Stores that already keep a Google Shopping feed can start selling through the chat agent without building a separate catalog, and shoppers are less likely to be quoted an old price.

Agent capability

The AI Agent Action in HighLevel workflows can now search the web in the middle of a run, switched on from a new Built in tools section. The agent decides when a task needs facts it does not have, and builders can also paste trusted page links into its instructions for it to read first. Each search is billed on top of the action, at 1 cent with OpenAI and Claude models and 1.4 cents with Gemini models.

Agencies can let follow up agents answer from current prices and programs instead of stale instructions, but should price the per search charge into what they resell.

Human approval / guardrails

Version 4.26.2 of the Inngest TypeScript SDK adds experimental command shims for Sandboxes. A shim makes a named program inside the sandbox call a handler in the customer's own app first, and that handler can do the privileged work itself, run the real program, or refuse it. Credentials never have to enter the sandbox, and commands that use shims can now run for longer than five minutes.

Teams running agents in Inngest Sandboxes can block a risky command, such as applying infrastructure changes, and keep secrets in their app rather than in the agent's environment, though the API is still experimental.

MCP / tool calling / API

Admins can now connect their own MCP servers, such as Notion, Linear, GitHub or Datadog, to Monte Carlo. Bayes, Monte Carlo's AI teammate, then calls those servers' read only tools in chat and while troubleshooting alerts, pulling in related incidents, tickets, code changes and logs as evidence. Servers can use no auth, an API key, shared OAuth credentials or a per user OAuth sign in, though alert troubleshooting works only with shared credentials for now.

Data teams get root cause analysis that draws on the tickets, code and runbooks they already keep elsewhere, and with per user sign in Bayes sees only what each person can see.

Agent capability

OpenCode v2.2.1 adds a /goal command in the terminal app and Desktop that keeps a session working toward a stated objective until a reviewer confirms it is met. Progress is reviewed automatically along the way, and the user can pause, resume or clear the goal at any point. Plugins also get an experimental hook to carry on work before a session goes idle.

Developers can hand the coding agent a whole objective and leave it running, with the review check deciding when it is done instead of a person prompting each step.

Integrations

Teams can now add Pally to a Slack workspace and chat with the assistant in direct messages and in channels. Drafts Pally writes can be approved right in the channel where the work is happening.

Teams that live in Slack get the assistant and its approval step in one place, so a drafted reply or action does not sit waiting while someone opens another app.

Integrations

Parloa Marketplace brings ready made connectors for systems such as Salesforce, ServiceNow, Zendesk and Microsoft Dynamics 365 CRM into the agent builder, alongside Solutions vetted by Parloa for security and AI safety. A Solution packages a whole service workflow, its questions, actions, rules and routing, and is added to an agent through a guided setup and then adapted before go live. Central teams can also share an agent with other regions and brands and see which version runs where.

Each new customer service use case can start from a tested workflow and existing integrations instead of another round of IT reviews and flow design, and large companies keep agent quality consistent across business units.

Agent capability

Version 3.0 of the Sema4.ai platform adds Verified Queries and Verified Logic, approved data lookups, calculations and checks that agents reuse instead of having the model write them again on every run. Each one is reviewed, versioned and tied to the work it runs in. Sema4.ai measures 80 percent fewer model tokens for Verified Queries and 50 percent fewer for Verified Logic compared with free form generation.

Repeat back office work such as reconciliations runs the same way each time at a lower model bill, and each step can be traced for a control review.

Integrations

WisdomAI now answers data questions inside Microsoft Teams, in a one to one chat or when someone mentions it in a channel. The answer lands in the thread where people are talking, and agent results and dashboard summaries can be delivered to Teams too. Answers follow the same domains, permissions and row level security as the web app.

Companies that run on Teams rather than Slack can bring governed analytics into their daily conversations without setting up a second permission model.

Agent capability

Gaia, the Gorgias assistant, can now turn any answer into a file: a PDF recap, a CSV of the underlying data, or an interactive HTML report to click through and share. It works in any Gaia chat and in Routines, so a standing request such as a Monday recap of last week's support and AI Agent performance arrives as a file on schedule.

Support leaders can send stakeholders recurring reports built from the same numbers as Gorgias reporting, without exporting from dashboards by hand.

Agent capability

Ontological Legal Research is live in the United States, pairing collected US primary law with a map of how legal sources connect. The map links statutes to the cases that interpret them, overruled cases to the decisions that overruled them, and rules to their exceptions, so the Legora Agent sees later treatment whenever it uses a source. Legora reports that the structure makes the Agent 40% faster on research runs.

Research agents that miss an overruling decision or an exception cite bad law, so a prebuilt map of how sources relate goes at the main accuracy risk in AI legal research.

Agent capability

Phonic-1 is Phonic's own speech to speech model for voice agents, available to customers now. It can call tools and act while the caller is still talking, so the answer is ready when they stop instead of after a pause. It is trained on more than 50 languages and can switch language mid conversation while keeping the same voice.

Teams building voice agents can replace a stack of separate transcription, reasoning and voice models with one model that hears tone and pauses the transcript would drop. Phonic-1 scored 83.2 on Sierra's multilingual τ³-Voice benchmark.

Agent capability

Primebox now writes AI reply drafts for email and LinkedIn conversations that come from Multichannel sequences. The AI reads the full conversation history and puts a draft in the reply editor for the rep to review, edit and send. Each email draft uses 3 personalization credits, rephrasing is free, and conversations already handled by automatic AI replies do not offer the option.

Reps answer interested leads faster across both channels while a person still approves every message, which suits teams not ready to let AI reply on its own.

Human approval / guardrails

Admins can now require approval before a Playbook, Writer's saved agentic workflow, becomes available to anyone beyond its owner and collaborators. Each new version also has to be approved before it reaches other users. Owners and collaborators can keep building and testing while a version waits for review.

IT and compliance teams get a review step before any automated workflow, or a change to one, spreads to colleagues, which matters as business users build more of them on their own.

Security / enterprise

Zep accounts can now have their Members and Groups provisioned from an identity provider over SCIM, with each identity provider group mapped to a Zep role. SCIM provisioning is in beta on the Enterprise plan, and Zep turns it on for each Enterprise account on request.

Security teams can grant and remove access to the agent memory platform from the identity provider they already run, which speeds up onboarding, offboarding and access reviews.

Workflow orchestration

Workflows in Atomicwork can now start in two new ways. Each webhook workflow gets its own endpoint, so any HTTP call from an outside system starts it and the data sent with the call can be used inside the workflow. Scheduled workflows run on their own like a cron job, with no ticket, request or user record attached.

IT teams can have monitoring alerts, business apps and scripts kick off service automations directly, and nightly cleanups or recurring reports no longer need a ticket to hang on.

Human approval / guardrails

ControlUp Connect becomes Pulse AI Resolve, an agent on each employee's device that diagnoses and fixes problems and answers questions so no ticket is needed. Admins now decide which fixes it may run on its own and which need a person's approval, write guardrails for what it must never do, and can hold some devices to fixes their team has already reviewed. Resolve can also act in other systems such as identity or security tools through MCP connectors, with approval set per action.

IT teams get a dial for how much the remediation agent does alone, device by device and action by action. It is a public tech preview, sold as an extra on top of Desktops Advanced or ControlUp ONE licenses.

Agent capability

The AI Analyst in Trends + Insights now answers natural language questions with reports, charts, comparisons and text documents, and shows the supporting evidence behind each answer. Users can explore themes with follow up questions, compare groups or time periods, add charts to dashboards and schedule recurring report digests.

Contact center leaders can question their conversation data in plain language, and seeing how conversations were sampled and analyzed gives them a way to judge each result.

MCP / tool calling / API

Bifrost v2.4.0 now discovers the prompts, resources and resource templates that upstream MCP servers offer and serves them through its MCP gateway alongside tools. Each server gets allow lists for what it exposes, virtual keys narrow access the same way they do for tools, and every read is logged. The release also supports MCP protocol version 2026-07-28, and tool results now keep images, audio and structured content instead of flattening to text.

Platform teams that route agent traffic through Bifrost can govern MCP prompts and resources with the same keys and logs they use for tools. Browser clients from origins that are not allowed now get a 403, so those should be checked before upgrading.

Workflow orchestration

Continuous Coverage, now available to all customers, has Sybil test what changed in an application instead of waiting for the next pentest. As often as daily, it reads code merged in GitHub, re crawls the live app for new endpoints, or both, then ranks the changes by risk and rolls them into one test on the customer's schedule. Validated findings go to Linear, Jira or GitHub Issues, and for white box apps Sybil can open a pull request with the fix.

Security teams can keep pace with how fast code ships, with or without sharing code access. Each test waits for approval by default, can be switched to run automatically, and every cycle is logged for audit.

Agent capability

The Misconfiguration Researcher is a new capability in Terra's automated pentesting that starts from the technologies actually present in an application's stack rather than a fixed list of known bad settings. It researches what each product can do and which configuration choices control those features, then passes any dangerous combination it finds to the validation phase as a hypothesis to test. One example is a FreeSWITCH deployment whose exposed interfaces open a path to running commands on the host with no flaw in the application's own code.

Security teams get findings from third party components and deployment settings, a class of risk that code analysis alone does not see.

MCP / tool calling / API

Arcade now supports MCP Apps, so a tool can show an interactive interface inside the chat instead of a block of text. Drafts from Gmail, Outlook Mail and Slack open as editable messages that can be sent from the chat, Slack messages can be scheduled, and files can be uploaded and downloaded. Each app shows only the actions the user's permissions allow, so a toolset with no send tool shows no Send button.

People can fix and send an agent's draft without leaving the chat, while the security team's policy is still checked when the interface loads and again on every click.

Memory / state

Facts that contradict each other in a Cognee dataset can now be listed and settled through the API and SDK in v1.6.6, by picking one fact with an optional reason. The decision is recorded and kept by later automatic reviews, and retrieval uses the chosen fact while the rejected ones stay stored. The improve step now also reviews conflicting facts by default.

Teams that give agents long term memory can correct contradictions on the record instead of leaving agents to answer from facts that disagree.

Agent capability

Think-Live is a Deepgram LLM built only to run the live conversation in a voice agent, deciding when to listen, wait or respond so the agent stops talking over callers and leaving dead air. It is generally available in the Deepgram Voice Agent API and can be paired with the model a team already uses. Think-Fast, its companion for following business rules and calling tools mid call, is in early access on the Deepgram LLM API.

Teams can fix turn taking in their voice agents without replacing the model that handles their business logic, which tackles the most common caller complaint about voice bots.

Agent capability

Podcasts are now part of the Exa index: 3.2 million episodes from 119,000 shows, each with a full transcript, across all endpoints. Results include hosts and guests, publish date and a transcript with speaker labels, so an agent can find what a named person said on recent episodes. New episodes appear soon after they air.

Research, sales and recruiting agents can draw on what executives, investors and candidates say out loud, often before they write it down, at no extra cost to current Exa usage.

Integrations

Antigravity App v2.28.2 brings partner plugins from Canva, Figma, Linear, Notion, Stripe, Supabase and Vercel to the Marketplace, with a new Featured spotlight. Plugins that connect to an account now take the user straight to sign in when added, and get an Accounts tab for managing connections. When a plugin would help with a task, the agent can also find it in the Marketplace and suggest it in the conversation.

Developers can hook the agent up to the design, planning, payments and hosting tools they already use with a sign in instead of manual setup. Plugin sign in and agent suggestions are not yet available on business and enterprise accounts.

Human approval / guardrails

Approve for me arrives in interactive V3 sessions with Kiro CLI 2.36.0. Kiro approves routine actions on the user's behalf, blocks risky ones and asks when it is unsure, marking each action it approves as Approved for you. Existing deny, ask and allow rules still decide first, and the setting is saved per workspace so it stays with that repository.

Developers stop getting interrupted by routine approval prompts while their explicit rules keep priority, which widens what the agent does without asking. It is not offered in every session yet, and Kiro keeps asking where it is unavailable.

Human approval / guardrails

Apps built on Builder.io's Agent-Native framework can now take approval for a sensitive MCP call inside the app when the calling MCP client cannot show an approval prompt. In version 0.221.0 the client gets a link to the app's review page, where only the same signed in user and organization can decide, and an approved call runs once within ten minutes. The audit log records these approvals as their own signal, mcp-review.

Teams exposing app actions to Claude, ChatGPT and other MCP clients keep a human approval gate even when a client has no prompt of its own, where before such calls simply failed.

Integrations

Agent output in Fabrix.ai can now be shared to Slack, Microsoft Teams or email, or downloaded, starting with RDAF release 8.2.3. The same release adds Azure Insights bots that collect amortized cost, savings plan use and Azure OpenAI token counts, plus an AWS bot that returns EC2 hourly prices.

Findings from Fabrix.ai agents can reach responders in the chat and email tools they already watch, instead of staying inside the platform until someone copies them out.

MCP / tool calling / API

A hosted MCP server now puts a company's Maven agent inside ChatGPT, Claude, Muse, Cursor and any other client that supports remote MCP, through a single tool, ask_maven. Each agent gets one URL, and a conversation ID in every answer lets follow up questions continue the same thread. Access requires the agent's API key or a sign in through any OpenID Connect or OAuth 2.0 provider, and the signed in user's token passes to agent actions without the model seeing it.

Customers can reach the same governed support agent from the assistants they already use, and actions run under each signed in user's own permissions.

Agent capability

Studio Pro 11.16 adds Plan Mode to Maia, the agent built into the Mendix IDE, so it drafts an implementation plan from a prompt that developers can review and refine before any build starts. Studio Pro now includes Playwright, which Maia uses to check the running app in its PDCA cycle. The Changes pane also shows which elements and properties Maia changed inside documents.

Teams can review the agent's plan before it touches the app and see it test the result, so changes the agent makes are supervised rather than accepted blind.

Integrations

Clinical guidelines, guidance and implementation resources from the American Society of Addiction Medicine (ASAM) now appear in OpenEvidence answers. The guidance surfaces automatically when it is clinically relevant, next to peer reviewed literature, with citations and a direct link back to the source. It covers the care of patients with substance use disorders and addiction.

Clinicians treating addiction get the specialty society's own guidance in the same answer as the research, which matters to primary care and behavioral health groups weighing answer coverage.

Agent capability

Pro users can now call patients and run video visits from a new scribe in the Scribeberry web and mobile apps, and after the call ends one tap generates the note from its transcript. Patients join a video visit from a texted browser link, with no app or account. Calls reach Canadian and US numbers and show the clinic's caller ID.

Clinics that offer virtual care can document phone and video visits in the same tool as in person visits, without a separate telehealth product.

Integrations

Sites protected by hCaptcha can now recognize Skyfire KYA tokens, the signed credentials that tell a site which agent is calling, whom it represents and what it may do. hCaptcha validates the credentials in standard request headers and feeds the verified identity into its policy engine and agent scoring, so site owners can admit KYA verified agents on chosen pages, workflows or the whole site. No separate Skyfire verification service is needed.

Agents carrying KYA credentials face one fewer bot wall on hCaptcha protected sites, and merchants can let them in through the security tool they already run instead of building their own agent checks.

Integrations

Argus, Cosmon's CAD agent, now works in CATIA V6, running it alongside CATIA V5 as separate application instances. Version 0.1.35 also adds Creo 9 and 10 to the Creo 11, 12 and 13 releases it already handled. When a user presses Stop, the agent now keeps the work in progress instead of throwing it away.

Engineering teams on CATIA V6 or older Creo versions can put the agent to work in the tools they already run, without waiting on a CAD upgrade.

Agent capability

Lindy Apps lets a team ask Lindy, in Slack or chat, to build an internal tool such as a deal room, spend tracker or approval board. Each app saves its own data, gets its own link and can be rolled back to an earlier version. Apps read from and write back to connected tools, so moving a deal in the app updates it in Salesforce, with a separate permission for each tool.

Business teams get working tools on top of the systems they already use without waiting on engineering, and removing someone from the workspace cuts their access to every app at once.

Agent capability

Voice AI agents built on ConversationRelay can now change most session and language settings in the middle of a call, such as speech providers, voices, timeouts and interruption rules, without dropping the caller. With application reconnect turned on, Twilio holds up to 500 events while the connection to the agent's app is down and replays them once it is back. Echo filtering stops an agent from reacting to its own voice coming back through the caller's microphone, and ElevenLabs v4 Turbo voices are supported, all now generally available.

Teams running phone agents can switch language for a caller mid conversation and redeploy or autoscale their app without cutting live calls off.

Security / enterprise

Python SDK 7.0.0 locks down the local Band MCP server that agents use: every request now needs a valid bearer token, and requests without one are refused. The built in Claude SDK, OpenCode, Letta and ACP integrations create and send the token on their own, while teams that wired up their own MCP clients must add it. The same release adds a native adapter that brings OMP agents into Band chat rooms.

Security teams get an authenticated tool endpoint for agents on shared machines, though the token does not guard against a process running as the same OS user.

Agent capability

Expressive conversation arrives in alpha in Pipecat 1.15.0: the LLM writes expression tags into its reply and the voice service performs them instead of reading them aloud. The LLM learns which tags its voice service supports, with adapters for Cartesia, ElevenLabs, Fish Audio, MiniMax and others. A perception feature also tells the LLM how the user spoke each turn, and new Pipecat Tone models read the user's tone and tag plain replies.

Voice agent builders can make agents that react to how a caller sounds and answer with matching delivery. It is an alpha, and the tag syntax and APIs are likely to change between releases.

Integrations

Release v2026.10.28 lets teams connect Sage Intacct to Pluvo directly as a data source. The same release can take a whole Google Drive folder and turn its supported files into tables automatically. The Assistant's model list is also replaced by a Thinking control with four levels, from Quick to Deep.

Finance teams that run Sage Intacct as their ERP can now point Pluvo's agents and reports at their ledger data, which removes a common blocker to trying it.

Agent capability

Prism is Rime's new speech to speech model for voice agents, available through a cloud WebSocket API and Rime's SDKs for Python, Node.js, Go and Rust. It handles turn taking and interruptions on its own, and calls a team's tools during a call while speaking filler lines as they run. Instructions and tools can change mid call, and Prism streams reply text alongside the audio for captions and logs, in English with four voices.

Teams building phone agents get one model that listens, acts and speaks, and because it uses OpenAI Realtime event names, an existing Realtime client can move over with a few changes.

Agent capability

The AI Receptionist now tells apart ten kinds of caller instead of four, adding six built for law firms: Other Law Office, Government Agency, Case-Related Third Party, Opposing Party, Media/Press Inquiry and Incarcerated Caller. Firms set each type up in the dashboard, choosing whether it goes to the AI or a live agent by warm transfer, which questions are asked, and whether the call ends in a message, a booking or a transfer.

Law firms can have courts, adjusters and opposing parties triaged by their own rules from day one, with no support ticket or custom build. It is on by default for every AI Receptionist customer, including the Free Plan.

Workflow orchestration

Voice Step 2.0 lets a phone call step in a Spoki automation branch on how the call ended: no answer, voicemail, goal reached, call back later, not interested or everything else, plus a Failed branch that is always present. Each outcome a team switches on becomes its own path for the next steps. New Voice steps use 2.0 by default, existing ones keep working as before, and a toggle returns a step to the classic version.

Teams that call customers from WhatsApp journeys can send each result to the right follow up, such as a WhatsApp message, a callback task or a CRM update, instead of ending the flow after the call.

Observability / auditability

Self Healing Workflows start on their own when a monitored evaluation score falls below its threshold. StackAI proposes a fix with a plain language explanation tied to the failing examples, applies it to a draft rather than production, and reruns the evaluation so the team can compare scores before and after. The change goes live only after the team approves it, and the same loop can be run by hand from any experiment.

Quality drift in live agents gets a proposed repair instead of just an alert, while a person still signs off on every change that reaches production.

MCP / tool calling / API

A new Conversations API, in beta, puts the same Assistant that sales reps use in Actively inside a customer's own tools. Each conversation keeps its own history, replies stream back as they are written, dropped connections pick up where they left off, and a reply can be stopped partway. A conversation can be tied to one account for an account page or left open for questions across a rep's whole book, and it also shows up in that rep's Actively sidebar.

Revenue teams can add account aware chat to internal tools and account pages without building their own chat layer, and reps see the same thread wherever they pick it up.

Human approval / guardrails

App Rules in Gumloop 13.2.0 now run on a decisions model. Admins describe in plain language what a tool call must not do, and the model judges each call against that rule to block it, tag it, or hold it for human approval. The same release lets users approve a background subagent's requests from the main conversation in chat, Slack, iMessage or email.

Admins can put a human sign off on risky agent actions by writing a sentence instead of coding a check, which narrows what agents may do on their own.

MCP / tool calling / API

Cross App Access now lets Okta admins connect AI agents to third party MCP servers, with full visibility and control over what agents do on a user's behalf and no consent prompts for users at run time. In the same release, MCP servers registered in Okta can act as OAuth clients to the systems behind them, with a separate connection and scopes for each system that can be revoked at any time. Both are generally available in Preview orgs with release 2026.44.0.

Security teams decide centrally which agents reach which outside MCP servers, and downstream systems log the actual user instead of a shared service account.

Security / enterprise

Enterprise admins can now add private npm registries to Retool, and builders can then ask the app builder agent to install packages from them, such as a design system hosted on the company's own servers. A registry can be the default for all packages or take whole npm scopes, and can sit behind a bearer token or basic auth. Teams that run their own registry can also set rules on what it allows, for example a minimum package age.

Security teams decide which code the agent may pull into the apps it builds, which answers the supply chain question raised by apps written with AI. It is a public beta for Enterprise plans, live on cloud and in 4.88 Edge.

Agent capability

Responders can now mention @Rootly in Slack to find who is on call for any team, now or when an incident started, and to create a schedule override or post a shift for a teammate to pick up. The same agent creates an incident's retrospective once the user confirms the template, and answers how to questions from Rootly's documentation. It acts with the user's existing Rootly permissions and replies in the thread so the channel sees what was done.

On call chores that used to mean leaving the incident channel now happen in it, and the agent cannot do more than the person asking is already allowed to do.

MCP / tool calling / API

Shopify storefronts now support the Universal Commerce Protocol permalink capability on a new /buy route, starting with UCP version 2026-08-25. A single link can add items by variant, prefill the buyer's email, country and language, apply discount codes, signal a Shop Pay preference and send the buyer straight to checkout. The same link format works on any shop that supports UCP.

Agent builders can hand a shopper off to checkout with one link format across shops instead of writing cart logic per store. Selling plans, line item properties and cart notes still need the older /cart links for now.

Agent capability

Sidekick can now spot a storefront improvement, build a draft of the change, and compare it with the current storefront using simulated shoppers driven by AI. It picks changes that can be tested as theme changes and are expected to affect add to cart behavior. The merchant reviews each suggestion before deciding whether to apply it.

Merchants get a staged draft with evidence behind it rather than an untested idea, though the evidence comes from simulated shoppers and not from a live A/B test.

Human approval / guardrails

Compliance Console, built into the Horizon platform, sets the rules Sierra agents follow when they reach out to customers by phone, text or email. Companies choose what the agent discloses on each channel, the hours it may call or text based on the recipient's local time, regional holidays when outreach is held, separate frequency limits for marketing and transactional contact, and how opt outs are honored. Dashboards in Insights show how many outbound calls and messages were allowed, deferred, blocked or held for review.

Lenders, insurers and collections teams can let agents run outreach over weeks or months while keeping each contact inside the local rules, with a record of what the agent was and was not allowed to send.

Workflow orchestration

Siit workflows can now run device management actions on phones, tablets and computers managed in Iru, JumpCloud and Microsoft Intune. A workflow can clear or reset the passcode on the device an employee picked in their request, clear Activation Lock on an Apple device through JumpCloud so it can be wiped and reassigned, or restart and shut down a device through Iru and JumpCloud.

Lockouts and device handoffs can be resolved from the employee's request without an admin opening the MDM console. The actions need the Standard or Pro plan with one of those MDMs connected.

Pricing / packaging

Workspace Owners and Admins can now cap the credits a single worker, or all of one member's workers, may use each billing period, from a new Usage limits page that also shows what each has used so far. A worker that hits its cap stops until the limit is raised or removed, and Owners, Admins and the worker's manager get an email. In the same release, a recurring task that fails or is blocked three runs in a row is paused and its manager is told why.

Finance and IT get a hard ceiling on agent spend per worker or per person, so one runaway task cannot drain a shared credit pool.

Agent capability

Autopilot Chat returns to Orchestrator and now works directly on the customer's tenant instead of only answering questions. It can manage processes, jobs, queues, assets, triggers, machines and folders, look into failed jobs and access issues, and help plan, build and deploy automations. Admins pick an approval mode, from asking before every operation to running with no approval prompts, and can connect MCP servers and switch tools and skills on or off.

Operations teams get an assistant that changes the live automation estate, so the approval mode becomes a policy decision about how much it may do alone.

Agent capability

V3 projects can now route user messages with Voiceflow Classify, an intent classifier that uses Voiceflow's own decision model to pick the intent that best matches each message. It is faster than classifying with a large language model, needs no training, and returns no match instead of guessing when nothing clearly fits. When it cannot classify a message, for example with more than 254 intents in scope, the agent falls back to the LLM method for that message.

Teams running intent based agents get quicker routing with no confidence threshold to tune. It does not apply to V4 projects, and existing projects keep their current setting until someone switches it.

Workflow orchestration

Durable workflows in DBOS can now run on sharded Postgres, with each workflow's state, step checkpoints and messages kept together on one shard. In a benchmark of simple no op workflows, throughput scaled in a straight line from 20K workflows per second on one Postgres server to 1M per second on 50 shards, using PgDog as the router. The benchmark code is open source.

Teams running large volumes of agents or background jobs can grow past a single Postgres server by adding shards, without bringing in a separate orchestration system.

MCP / tool calling / API

Connectors, GitHub verified MCP servers, are rolling out in public preview on all Copilot plans. A developer picks a service and signs in once, with no server to find, nothing to install and no credentials to manage, and 26 services are available at launch, including Slack, Figma, Notion, Sentry, Linear and Vercel. A connected service then works across the Copilot app, Copilot CLI, VS Code, Copilot in the cloud and GitHub Mobile.

Copilot can draw on context from the tools a team already uses without each developer setting up MCP servers by hand. On Business and Enterprise plans an admin must turn on the Editor preview features policy first.

Human approval / guardrails

Action approvals let the incident.io agent use tools that change things, such as turning off a feature flag, once a person has signed off. Admins can set any write or destructive tool to Allowed with confirmation, and the agent then proposes the exact call and runs it only after the requester approves, as that person. Every proposal, decision and execution lands in the audit log, and approvals work in Slack, the dashboard and mobile.

Incident teams can let the agent fix production problems, not just report on them, with a named human approving each change and a record for the postmortem. Approvals for actions requested over MCP are not available yet.

Agent capability

Realtime TTS-2 has a new opt in Reasoning mode that decides how a line should sound before speaking it. The model reads the text and any context sent with it, writes a short delivery direction such as warm and delighted, then performs it, and the direction comes back in the API response. It works on synthesis, streaming and async requests today, while WebSocket streaming, the Realtime API and batch synthesis are still to come.

Teams building voice agents and companion apps get delivery that fits the moment without keeping a list of emotion tags and prompt rules, though live voice sessions on the Realtime API cannot use it yet.

Agent capability

Version 1.8.3 of the Kore.ai agent platform lets a voice agent send the caller a WhatsApp message with tappable options, such as appointment slots, while the call is still live. The message goes through the project's WhatsApp connection on Meta Cloud or Infobip. The caller's tap or typed reply comes back into the call as input, on both pipeline and realtime voice.

Callers can pick from a long list of options on their phone screen instead of hearing them read out, which shortens calls that end in a booking or a choice.

Workflow orchestration

In Moderne CLI v4.12.0, a parent agent session can see, steer and stop the agent sessions it starts, and a session that starts others is told how to drive them. The release adds the Pi coding agent to the agents the CLI can run and configure. A recipe can now also run in Moderne across the repositories a changeset found.

Teams rolling out code changes across many repositories with several coding agents can supervise them from one lead session instead of watching each one separately.

MCP / tool calling / API

Default provider routing on OpenRouter no longer picks mainly on price. The new router tracks each endpoint's real price, latency, throughput and error rate, chooses the best mix of cost and speed for each request, and moves traffic away from an endpoint while it has trouble. In OpenRouter's measurements, time to first token improved 16.9%, throughput rose 16.3% and capacity errors fell 46%, with costs up 6.0%.

Agents that make many model calls in a row get faster, steadier responses without any code change. Teams that want the lowest price on every request can still set an explicit price sort or pin providers.

Human approval / guardrails

Policy Studio turns company policies into rules that govern Aera's decisions as they are made. Teams upload a policy document or describe a rule in plain language, Aera drafts the rules for a person to approve, and each rule can be tested against live decisions before it goes on to block a decision, require approval or raise a warning. A new Decision Ledger keeps a traceable record of every decision, including those sent in from other systems through Aera's API, with the context, reasoning, policies evaluated, actions and outcome.

Compliance and operations leaders get written policy enforced on automated decisions, with an approval gate where it is needed and a record of how each decision was governed for audit and review.

Agent capability

Analog chip design now sits alongside digital on the ChipAgents platform, with agents covering topology and schematic design, simulation, behavioral modeling, layout generation and physical verification. Both run on IC-star, an engine that keeps the customer's design knowledge in a persistent layer, plans and coordinates agents and engineering tools over long workflows, and uses physics aware models to predict timing, PPA, DRC and DFT. A model navigator routes each task across AI models by quality, time and cost, and the platform works with existing engineering tools on premises or in the cloud.

Semiconductor teams can put digital and analog flows on one agent platform that works out the next step itself while engineers keep control of requirements and signoff, without replacing their design and verification tools.

Agent capability

A new auto mode in Claygent picks a model for each research task, and it is now on by default for new Claygent columns. Choosing the model takes under a second, and users can still pick a specific model whenever they want. The router is an open weight model Clay post trained with Applied Compute, running on Applied Compute's AC2 platform.

Clay puts the saving at 47% against sending every task to the most capable model, so teams running research at volume can cut spend without tuning model choice column by column.

Integrations

New connectors for n8n, lemlist and Livestorm are now in the Connector library under Workspace Settings. An admin sets up n8n first, and then each user connects their own account. Background skills also now send a bell notification and an email when a connector they depend on has expired.

More of a team's existing tools can feed Day AI without custom work, and the reconnect alerts keep scheduled skills from failing silently when a connection lapses.

Agent capability

Media Studio is a video agent inside Genspark that plans a piece, writes the script and storyboard, then directs image, video, audio and 3D generation models through quality checks, editing and compositing. It agrees on concept, script, storyboard and budget with the user before rendering anything, and brings each key checkpoint back to the conversation. Finished assets sit on one canvas with a layer editor, a multi track timeline and a 3D director stage that people and agents can both use.

Marketing and content teams get one agent that carries a video from a one sentence brief to a finished cut, from a 15 second ad to a 10 minute short, with a budget sign off before any rendering.

Integrations

A NodeZero integration for Palo Alto Networks Cortex XSOAR is now on the Cortex Marketplace, certified by Palo Alto Networks. XSOAR takes in open high and critical NodeZero findings as incidents, complete with affected asset data, exploitation evidence and remediation guidance. After a fix, teams retest with NodeZero 1-Click Verify, and findings that are no longer exploitable stop coming back as open incidents.

Teams that already run remediation playbooks in XSOAR can work from findings that testing proved exploitable, then confirm each fix worked without leaving the loop.

Deployment / data residency

Muse, Meta's personal agent, is now available in Singapore on iOS, Android and muse.ai. Local connector partners cover shopping, groceries, dining and travel through Shopee, Lazada, GrabMart, Grab Dine Out, Singapore Airlines, Agoda, Klook, Trip.com and ShopBack, alongside Gmail, Google Calendar and Instagram. Some connector features roll out in stages after launch.

Muse enters its new market with connectors to the shopping and travel services people there already use, and it still checks with the user before sensitive steps such as sending an email or making a purchase.

Memory / state

Sandboxes can now save their filesystem automatically on exit with Exit Snapshots, available in beta. A snapshot is taken on graceful and ungraceful exits alike, including terminate(), entrypoint completion, idle timeout, lifetime expiry and out of memory errors.

Teams running coding agents in Sandboxes can resume an agent after an idle timeout or recover work from a Sandbox that died, without writing their own snapshot logic.

Observability / auditability

An Agents Usage dashboard in nexos.ai lists every agent with its usage, credit cost, model and integrations. Admins can track credit cost per agent for the billing period and in total, see which agents are used most or not at all, and filter, sort and export the view to CSV.

Companies running many agents can see which ones earn their spend and retire idle ones without building their own reports.

Agent capability

Recruiters can ask Nova to add a specific person to a search by LinkedIn link or by name, such as an applicant they want evaluated. Nova scores that person against the hiring brief the same way it scores anyone it finds, then adds them to the pipeline, where their contact details can be revealed or the profile sent to the ATS.

Applicants and people a recruiter already knows get judged on the same fit scale as sourced candidates, in the same pipeline, rather than being tracked outside the tool.

Agent capability

Work Mode, now available to all paid Otter users, builds a brief before each meeting from past conversations with the customer, open action items, what worked for teammates on similar calls and the latest files from connected apps such as Google Drive. After the call it produces finished work, such as a proposal or contract as a PDF, an offer letter, a candidate scorecard, CRM updates and follow up emails, each ready for the user to review and send. Live Assist, private coaching tips shown inside Zoom, Teams and Google Meet calls, is live for enterprise customers and reaches all paid users over the next few weeks.

Otter now does the follow up work after a meeting, not just the notes, which matters to sales and recruiting teams comparing meeting agents on what happens after the call.

Observability / auditability

A new Replicant app puts calls and chats handled by AI agents and by people on one Conversations page, with the same search, filters, scorecards and reporting. Scorecards now group items into categories and give each score a reason and a link to the supporting moment in the transcript, and customers can define their own Topics and Reasons. Ask AI, which answers plain language questions about conversation data, no longer has a 1,000 conversation cap, and the new app has been the default for all customers since October 28.

Contact center leaders can measure AI agents and human agents against the same scorecard in one place, and see where automation handles a call without resolving it.

Integrations

WhatsApp is now a native channel for SalesCloser's AI sales agents, switched on at each customer's request, and the agent answers leads there with the business's own services, pricing and hours. Threads land in a new Chats tab in the dashboard, with conversations that need a person pinned to the top. When a lead asks for a person, the agent pauses and emails the business a link to the conversation, then resumes on one click or takes the chat back if no one replies in time.

Businesses whose leads prefer WhatsApp can put the same agent there, with a set takeover and handback routine instead of an agent improvising when someone asks for a person.

Agent capability

The AI Builder can now generate PDF Pages directly, and those pages support row level security and HIPAA protected data. Apps created through Vibe can also include buttons that generate PDFs.

Teams building regulated apps with the AI builder get document output without hand building PDF templates, and record level access rules still apply.

Agent capability

Login Steps lets a user describe an application's login flow in plain language, and Equixly's agents follow those steps to sign in before they run penetration tests in Projects. The same profile can carry an OAS security scheme and a description of what each user role may and may not do, which Equixly turns into an authorization matrix and tests against. Discovery has signed in this way since June, and that option now shares the Login Steps name.

Authenticated testing of apps with awkward logins no longer starts with long manual forms, and role limits written in plain words become tests that check users cannot go past them.

Observability / auditability

Decision model metrics in Galileo classify and score large volumes of text much faster and at far lower cost than metrics that use an LLM as the judge. This first release supports TypeSafe Jev decision models.

Teams that want to score every production trace from their agents, not just a sample, get a cheaper way to run evaluations at that volume.

Agent capability

Goal Oriented Agents are a new kind of Regal voice agent that builders point at an outcome instead of a script. Each goal node lists what the agent must collect or do, such as a caller's name or a scheduled callback, and the conditions that must hold first. A fast actor model carries the conversation while a director model checks progress after every turn and moves the call to the next stage, and existing customers get access through their Regal customer success manager.

Long intake and qualification calls, such as insurance claims or Medicare enrollment, can follow the caller's lead without asking for details twice, while the director still holds the agent to its guardrails on every turn.

Agent capability

Wonderchat Helpdesk puts email, live chat, WhatsApp, SMS and every handover from the AI agent to a person into one inbox, with two way sync to the team's own mailbox. The agent drafts or sends replies, sets priority and routes tickets, and can prepare account work such as a refund or plan change that a teammate approves before it runs. Claude, ChatGPT, Cursor or any MCP client can also work the queue, signing in with OAuth or a scoped API key.

Wonderchat goes from answering questions to running the whole support desk on every plan, with no per seat or per resolution fees, while account changes such as refunds still wait for a teammate's approval.

Agent capability

In version 5.18.18, users can describe a goal to the Operator, such as a Q4 ARR metric, and get back a card that opens the create dialog already filled in. Nothing is saved until the user clicks Create. A new Set up auto update option has the Operator walk the user through connecting a data source, so the numbers a goal reports each week stop being typed in by hand.

Goal setup and the weekly number update are the two chores that make OKR tools go stale, and the Operator now takes on both while a person still confirms each new goal.

Integrations

The FurtherAI Accelerator for PolicyCenter is now live on the Guidewire Marketplace and runs on Guidewire Cloud Platform. It takes a commercial submission from the broker's email, screens it for sanctions, litigation and adverse media, checks it against underwriting guidelines to recommend quote, refer or decline, and assigns an underwriter. It then creates the submission in PolicyCenter with the details filled in, and pauses for a person whenever clearance flags an issue.

Carriers and MGAs on Guidewire can move submissions from inbox to policy system without keying data twice, while keeping a human decision at the clearance step and a log of every action the agent takes.

Integrations

Runtime detections for AI agents from HiddenLayer now arrive in Splunk through its HTTP Event Collector and in Microsoft Sentinel through a Log Analytics table, and model scan results reach Splunk the same way. Each record carries requester, session and interaction IDs, a severity and a MITRE ATLAS tactic and technique, but no prompt text.

SOC teams can route and correlate AI agent events with the searches and analytics rules they already run, without managing a second queue. Agentic Runtime Security is rolling out tenant by tenant, so availability varies.

Agent capability

Itential AI Skills is an open source set of skill packs that lets teams build, run and administer the Itential Platform from Claude Code, Codex CLI, GitHub Copilot and Cursor through plain language requests. Builder Skills creates workflows, FlowAgents and other platform assets, Operator Skills runs and troubleshoots jobs and works on manual tasks awaiting a decision, and Admin Skills manages users, roles, SSO and integrations built from an OpenAPI specification.

Network and infrastructure teams can drive their automation platform from the AI coding tool each engineer already uses, while every action still runs under Itential's existing permissions and controls. An Itential environment is required.

Integrations

A remote MCP connector for Adobe Experience Manager now links Mark, Optimizely's agent platform, to a team's AEM environments, in beta. Agents can search and manage assets, content fragments, pages and launches from inside Mark. They can also check content against brand governance rules and troubleshoot CDN and replication issues.

Marketing teams that run AEM as their content system can put Optimizely's agents to work on that content without building their own integration. The connector is in beta.

Agent capability

Siena's customer service agent now creates returns and exchanges in Loop Returns inside the conversation, on live chat or email. It lists the items a customer can send back using the store's own return reasons and condition questions, reads back the refund, store credit or exchange, and files the return once the customer agrees. It then shares the prepaid label and tracking number in the chat, and can check a return's status or cancel it on request.

Brands on Loop Returns can have the agent finish a return from start to end on the channel the customer chose, instead of sending them off to a portal link.

Agent capability

AI Data Analyst, the conversational analyst built into the Abnormal platform, can now run one time or recurring scheduled reports, and any question can become an ongoing report with Schedule as Custom Report. A new Report Repository keeps every report and conversation with versioned downloads and full history. MSSPs and customers with several tenants can manage reporting for all of them from one view.

Security leaders get board reporting that keeps running on its own instead of a fresh query each month, and service providers can run it for every client tenant from one place.

Agent capability

In-Visit Orders picks up the orders an outpatient clinician talks through with a patient and lists them in the Heard in Visit Recording module in Epic. New orders appear as they are discussed, changed ones update in place and retracted ones are marked. The full set is staged in Epic for the clinician to review and sign before leaving the room.

Order entry moves from the end of clinic into the conversation itself, while every order still waits for the clinician's signature.

Security / enterprise

Organizations can now sign their staff in to Airia through identity providers that run on premises, including ADFS, Keycloak and PingFederate. The connection goes through the Airia Cloud Connector, so the identity provider never has to be exposed to the public internet.

Companies whose identity systems sit inside a private network can roll out Airia agents and tools without opening those systems up or moving to a cloud identity provider first.

Human approval / guardrails

Owners and Incident Responders can now stop a gamebook while it runs: the action in progress finishes, nothing further starts, and the gamebook ends with a canceled status and the reason given. Gamebooks that AI agents run directly against a user or device, with no incident attached, now appear in Gamebook history, where they can be approved, denied, canceled or retried. A new Gamebook Completed webhook reports how each run ended and the result of every action.

Security teams get a stop control over response actions already under way, and automated response now runs only through AI agents under the same approvals and audit as every other gamebook, since the per rule auto run setting is gone.

MCP / tool calling / API

Headless is a new plugin and MCP server that lets coding agents such as Claude Code, Codex and Cursor create and change Dataiku projects, including pipelines, models, dashboards and agents, by handing the build to Cobuild. It can also help move workflows built in SAS, Alteryx or Excel into Dataiku, and gives administrators tools that mainly inspect projects and failing jobs. It installs from the public Claude Code and Codex plugin marketplaces.

Developers can keep working in their coding agent while the result lands as a visual Dataiku project the wider team can review and govern, rather than a script only its author understands. It acts with the permissions of the Dataiku API key used to connect, and building needs Cobuild enabled.

Agent capability

One, the AI agent inside Factorial, can now build a full presentation from a single prompt using data that lives in Factorial, such as headcount, absenteeism and overtime per store. Users attach their company template and get back a .pptx with charts in that style, along with an explanation of where each number comes from. One can then revise slides, add metrics or plan for a scenario like Black Friday by estimating the extra staff each store needs.

HR and operations leaders can produce management committee decks from live HR data without exporting spreadsheets or formatting slides by hand.

Human approval / guardrails

Events & Hooks lets workspace admins and owners on Business and Enterprise plans connect Lovable to their own services. Events tell the customer's HTTPS endpoint when a project is created, published, moved to another workspace or deleted. Hooks let that endpoint allow or block an action first, for example checking a chat message for sensitive data before Lovable processes it.

Security teams can put their own policy check in front of what reaches the Lovable agent and keep an inventory of every project, without reviewing each one by hand.

Deployment / data residency

Version 3.0 is out as a public beta, a major release that clears out legacy features and makes security defaults stricter, with the stable version due in early December. Self hosted n8n now runs only on Docker, as npm installs are no longer supported, while Kubernetes and Helm setups are unaffected. Slack credentials now need a signing secret, unverified community packages are off by default, and the default Code node timeout drops from 5 minutes to 1.

Teams that installed n8n with npm, or that rely on removed nodes such as Function, Cron and Interval, have work to do before upgrading, and the move to 3.x cannot be rolled back. The built in Migration Report lists the affected workflows and fixes many of them automatically.

MCP / tool calling / API

Release 1.4.0 ships the holo command line tool, which runs H's browser and desktop agents from a terminal in interactive sessions where users send follow ups, pause the agent and correct it. A built in MCP server lets coding agents list H agents and start runs, and one install command adds it to 13 coding agent hosts or connects them to the hosted MCP server instead.

Developers can hand browser and desktop tasks from the coding agent they already use to H's agents without writing integration code, which widens where these agents get put to work.

MCP / tool calling / API

A Joule Studio command line tool now sets up Model Context Protocol connections, development skills and authorizations in one step, so a project opens in the developer's own IDE with its business requirements already in place. Agents built in Joule Studio can be deployed on the Joule Studio runtime, where they show up in SAP AI Agent Hub and in Joule Work conversations, and existing Joule Agents can be extended with instructions, MCP tools, APIs and workflow triggers through a governed pipeline. General availability is rolling out in phases.

SAP customers can build and extend agents in the tools their developers already use, while every change passes a review step before production and each agent stays registered where IT governs it.

Agent capability

Oak 1, the Speechmatics medical speech to text model that launched for batch jobs in mid October, now works for real time transcription too. It recognizes medical keywords better than Melia 1 at the same base accuracy, and developers switch to it with a single model setting when they open the streaming connection.

Teams building ambient scribes and clinical voice agents can now use the medical model during live consultations instead of only on recordings after the fact.

Security / enterprise

Workspaces can now connect Tenzai and run an AI penetration test on a published Base44 app from its Security page. The test attacks the live app the way a real attacker would, including pages shown after sign in, then emails a report of what it managed to do and how to fix it. Tests are billed to the customer's Tenzai account, not Base44 credits, and need the Builder plan or higher.

Teams shipping apps built by AI get an attack test of the live app from inside the platform, which covers risks a static security scan can miss.

MCP / tool calling / API

Evals datasets in Fiddler can now be edited through the REST API in Release 26.25, with every add, update or delete saved as a new numbered version that can still be read later. A write based on an out of date version is rejected and nothing changes. The Fiddler MCP server can now make the same edits in a single call and list a dataset's versions.

Agents and scripts can now maintain evaluation datasets without overwriting each other's edits, and every past version stays available when a test result needs explaining.

MCP / tool calling / API

Release 565 adds 57 remote server templates to Glean's MCP server library, among them Google Ads, 6sense, BambooHR, Kong Konnect, LinkedIn Hiring Pro, NetSuite QuickStart and Pipefy. Admins switch on the ones they want from the library. Most of the new templates sign in with per user OAuth.

Glean Assistant and agents can reach many more business systems without custom connector work, and per user sign in keeps each person's own access rights in place.

MCP / tool calling / API

The new watsonx Orchestrate AI Gateway is a single enforcement layer that discovers third party agents, models and MCP servers on platforms such as Amazon Bedrock, Microsoft and Google. Customers can now talk to those outside agents directly through watsonx Orchestrate chat and channels. The gateway applies the same observability, evaluation and supported policies to them as to agents built in Orchestrate.

Enterprises with agents spread across several clouds can find and govern them from one place, without rebuilding them or standardizing on one platform.

Workflow orchestration

Lexis+ with Protégé can now pull Lex Machina litigation analytics, Intelligize disclosures, CourtLink dockets and Law360 reporting into one prompt box, alongside a firm's own work product through document management connections. Lawyers can pick the sources themselves or let a governed orchestration layer choose the models, tools, skills and sources for the task. Sources stay visible and customer permissions still apply.

A litigation or deal team can ask one question and get research, analytics, dockets and its own prior work back together, each linked to its source. It is rolling out to US customers first, and which sources appear depends on the subscription.

MCP / tool calling / API

Admins can now create custom Model Context Protocol connectors in the Microsoft 365 admin center and deploy them across the organization. Microsoft 365 Copilot then draws on the company's own data through those connectors. Creation sits in the Copilot extensibility section, next to the federated MCP connectors admins could already deploy from the same center.

IT can bring a proprietary system into Copilot from one console and roll it out to everyone, instead of each team building its own integration.

MCP / tool calling / API

ChatGPT, Codex, Claude, Claude Code and Cursor can now connect to Databricks MCP servers, including the Genie One MCP server, with default OAuth client IDs. An account admin no longer has to create an OAuth app for each of these agents first. The feature is generally available.

Teams can point the external agents they already use at governed Databricks MCP servers without an admin setup step, while access still runs through OAuth.

Deployment / data residency

Each new conversation in Agent Canvas 1.34.0 can run on a Local, Worktree or Docker runtime, picked from Home whenever the Agent Server offers Docker. A new Runtime settings page sets the default runtime, the worktree options, and the memory, CPU and browser for the container.

Teams can put a risky agent session in its own container with set resource limits while keeping lighter work on the host, which narrows what a misbehaving agent can touch.

Human approval / guardrails

Agent Kill Switch is now generally available in Agent Fabric, the Salesforce control plane for governing agents across platforms, so a team can pull an agent offline from one place. Model Wallet, also generally available, sets hard spending limits that stop requests before costs pass the budget, and the FinOps Dashboard puts AI spending in one view. Cost Optimization Policies reduce token use through semantic caching, payload compression and response trimming.

Companies running agents on several platforms can stop a misbehaving agent and cap its spend without tracking down each platform owner. A kill switch that fires automatically from evaluation results is not due until January 2027.

Security / enterprise

Temporal Proxy, now in public preview, is a component customers run in their own environment that encrypts Workflow payloads before they reach Temporal Cloud. It works with AWS KMS, Google Cloud KMS, Azure Key Vault or a custom key service, so Temporal Cloud stores only ciphertext it cannot read. Making the proxy the required connection to Temporal Cloud means no application or Worker can skip encryption by mistake.

Security reviewers get one shared control to check, under keys the customer holds, instead of separate encryption code in every application, which matters as more agents send customer data through workflows. Temporal supports it for production use during the preview.

Human approval / guardrails

AI agent steps in Windmill flows gain tool approval and a question tool in version 1.850.0. An agent can now wait for a person to approve a tool call before it runs, and can put a question to a person in the middle of a run.

Teams pointing agents at production systems get a human checkpoint on individual tool calls, not only on whole flow steps, which limits what an agent does on its own.

Agent capability

Agent mode projects in Instabase AI Hub release 26.42 can now pick the model provider used for all of their LLM calls, or choose Auto to take the AI runtime's default. The choice is saved with the project and becomes the default for app versions published from it. It is an opt in feature on the Enterprise lite and Enterprise plans, live on multi tenant AI Hub from October 26 and on single tenant deployments from November 9.

Teams with rules about which model vendors may process their documents can fix the provider for each project instead of inheriting the runtime default.

Agent capability

The Retrieve Knowledge node in AI Agents (Cognigy) flows can now return raw content chunks from Knowledge Hub instead of only a finished generated answer. An AI agent can then reason over the source material and build its own response. The option is generally available in the 26.4 release cycle, which began on October 26.

Teams building agent flows choose between a ready made answer and the underlying source text, which matters when an agent has to combine knowledge with other steps before it replies.

Integrations

Sim 0.9.32 adds an Amazon EC2 integration. Agents and workflows can now inspect and control instances, create disk snapshots and manage resource tags, using an AWS access key with the required permissions. The same release lets teams export a knowledge base and import it into another workspace or deployment.

Agents built in Sim can now change cloud infrastructure, not only read data, so platform teams should limit the AWS key to the actions they want an agent to take.

Agent capability

Chat agents in Callers can now run actions before a chat starts and while it is under way, not only after it ends. All three kinds appear together in chat details, and the test chat panel now shows why a chat did not start, for example when an action set to run first fails.

Chat agents can gather context before their first reply and act during the conversation, which makes them more useful for tasks that depend on other systems.

Deployment / data residency

Self hosted deployment is now available for Actian AI Analyst, formerly Wobby, in a private cloud or on premises, where until now it ran only as managed SaaS. A bring your own LLM option connects the analyst to a customer's own OpenAI or Anthropic account, or to models reached through AWS Bedrock, Google Vertex AI, Azure AI or an enterprise AI gateway. The self hosted version keeps the same governed semantic layer and visible answer logic as the cloud one.

Teams whose data or approved model providers must stay inside their own environment can now put the analyst in front of business users, and model governance teams keep inference on endpoints they already run.

Integrations

A PitchBook connector in North gives agents PitchBook data on companies, deals, funds, investors and market benchmarks, alongside a company's own documents. Customers bring their own PitchBook license and sign in, and the connector keeps the permissions already set in PitchBook and North. Every figure North pulls from PitchBook carries a citation.

Corporate development, strategy and investor relations teams can draft memos and board updates on cited private market data without leaving North, provided they hold a PitchBook license and have single sign on turned on.

Integrations

Ray, Ocean's email investigation agent, can now be added from the Gemini Enterprise agent marketplace and connected to a customer's Ocean tenant. It joins over the open Agent2Agent protocol as an agent rather than a tool, so other agents hand it a task and Ray runs the full investigation, streams progress and asks a follow up question when a request is unclear. Other agents get Ray's verdict and evidence, never raw access to the email data.

Security teams on Gemini Enterprise can chain Ray into incident response, identity or finance agents without a custom integration, though for now Ray only answers and explains rather than taking action.

Observability / auditability

Generative dashboards are now generally available in Opik. A user describes the metric they want to see to Ollie, Opik's assistant, which writes a read only query over traces, spans, threads, experiments and feedback scores and offers several chart options to pick from. The chosen widget joins a dashboard and updates as new traces arrive, and Edit with Ollie reopens the conversation to change it later.

Teams can track something specific, such as the hallucination score for each agent version, without writing queries or designing charts, so it is easier to catch a release that quietly gets worse.

Observability / auditability

Simulations let teams test an AI Voice Agent before launch: an AI caller holds practice conversations with the real agent on its actual setup. Scenarios can be written by hand or generated by AI, and each conversation is scored against the team's success criteria with the full transcript and the reasoning behind the result. Outbound AI Voice Agent campaigns can also now run continuously with no contact cap, dialing new contacts until someone pauses them.

Teams can catch a voice agent's mistakes and regressions before customers hear them, which matters more now that one outbound campaign can keep calling thousands of contacts without a stop.

Agent capability

In release 2026.9, Spark, Nexthink's employee IT support agent, can manage tickets itself: checking status, reading and adding comments, and confirming resolution through configurable actions. Admins can set categorization and routing so that Spark's escalations follow their own ITSM processes. The same release lets Spark learn from previously resolved conversations and adds action cards that show what each remediation action changes before an admin enables it.

Service desks can let Spark carry a request through the ticket lifecycle rather than stopping at a handoff, with a clear view of each action's impact for security sign off. The ticket features need configuration and user permissions, and Spark may need an extra license.

Security / enterprise

ToolHive v0.57.0 adds authorization scoped to each backend in its virtual MCP server, so access rules can differ for each MCP server sitting behind the gateway. Token exchange now accepts authorization details, and upstream subject identifiers can be switched on as an option. Applications that embed ToolHive can also refuse refresh grants when the upstream credential has stopped working.

Security teams that route agents' tool calls through one MCP gateway can limit which backend each identity reaches, instead of granting the whole gateway at once.

Agent capability

Rehearsal, a new mode in Token AI, takes a goal in plain words, such as finding which AI agents have no owner, and carries it out in the user's own tab. It plans a route through the platform, opens the pages, sets the filters, highlights the matching results and writes up the answer. Reading is automatic, but it asks before saving anything, never fills in a secret field, and stops the moment the user takes over with a click.

Security teams get answers from a dense identity console without knowing where each one lives, and the limits on saving and secret entry make clear where the agent acts alone and where a person decides.

Deployment / data residency

Starting with v0.114.2, every Hatchet server image is also published in a FIPS 140-3 version, built against the validated Go Cryptographic Module. The same release lets teams pause a rate limit key so waiting tasks do not consume the limit, and lets a task choose when its next retry runs.

Teams in regulated or government settings that must run FIPS validated cryptography can self host Hatchet from published images instead of building and maintaining their own.

MCP / tool calling / API

Cloud coding agents in Hoplite can now use tools from MCP servers running on a developer's own computer, such as Paper or Figma. The hoplite mcp connect command links a local server to a cloud thread, handles browser sign in where supported and keeps credentials on the user's machine. A read only option turns off any tool that changes things.

Teams can give cloud agents access to design tools and other local tools without moving those credentials into the cloud, and can limit the agent to reading when that is all the task needs.

MCP / tool calling / API

Telli agents and workflows can now run code the customer supplies. A code tool calls the customer's own systems during a live call, and a code node runs code as a step in a workflow. Charlie, the built in assistant, can write and test the code, which can use the credentials and Salesforce or HubSpot connections already in the account.

Voice agents can reach systems that have no ready made integration while the caller is still on the line. Custom code is available on the Scale and Enterprise plans.

Human approval / guardrails

Adding a Vybe agent to a Slack channel now gives it write access by default, so it can post and reply there without extra setup. The same release lets an agent that needs a new app ask in chat or Slack and connect it once the user approves, instead of sending the user to settings.

Agents now speak in every channel they join unless someone limits them, so admins should review which channels agents are added to.

Browser/computer use

Clio Work can now fill in the submission form on a court's own electronic filing portal, using documents and matter details already stored in Clio. That reaches U.S. courts that Clio File was never connected to. The lawyer reviews and submits every filing personally.

Firms that file across many jurisdictions stop retyping matter data into each court's portal, while the final submit stays with a lawyer.

Agent capability

Deep scans in Devin Security now hunt for complex chained exploits across large codebases, using Claude Mythos and other frontier cyber models that enterprises reach through Devin without separate model access. Devin scores findings against a threat model built with the customer's team, then can document a false positive, apply an infrastructure mitigation such as a WAF rule, or fix the flaw in parallel sessions on separate virtual machines. Any proposal to relabel a finding needs human approval, and vulnerabilities on the KEV list are never relabeled.

Security teams get findings ranked by how exploitable they are in their own environment, plus an agent that tests and ships the fixes, which helps when the backlog outgrows engineering capacity.

Pricing / packaging

Agents built on ElevenAgents can now run supported LLMs on a fast service tier, which costs more per token in exchange for higher throughput and lower latency. Builders switch it on per agent or in workflow overrides, and the model list shows which tiers each model supports. The default stays standard, and the LLM usage calculator can estimate the cost difference before anyone switches.

Voice agents live or die on response time, so teams can now pay for speed only on the agents that need it and price that choice up front.

Agent capability

The CSM Adoption Agent is a Claude plugin that uses the Gainsight MCP connection to find customers with adoption gaps, drawing on usage, goals, success plans and account history. It drafts a personalized email for each customer and shows why that customer was picked. The CSM approves, edits or rejects each draft, and approved emails land in Gmail as drafts to send by hand.

Customer success teams get adoption outreach planned from their own Gainsight data, with a person signing off on every email before it goes out.

Workflow orchestration

SubAgents in Pydantic AI v2.62.0 can run a delegation as a task that the model can send to the background, steer, answer and cancel. Tool calls can also run as Temporal child workflows. A new BatchModelRequests capability adds support for provider batch APIs.

Teams building long running agent systems get supervised delegation and durable execution without writing their own task control, and batch calls can cut the cost of offline workloads.

MCP / tool calling / API

Skills committed to Team Context now reach the SageOx console, AI chat and a customer's Claude organization, each credited to the coworker who wrote it. With SageOx connected over MCP, up to 12 team skills show up as prompts in tools that support them. On Claude Team or Enterprise plans, an owner or admin can sync the team's plain prose skills into Claude for chat, Cowork and Claude Code.

Coworkers who never open a repository can use the same vetted skills as the engineers. The Claude sync copies the whole Team Context repository, which is worth a data review before turning it on.

Agent capability

The SonarQube Remediation Agent on SonarQube Cloud can now propose fixes for flaws found by the Hunter Agent, such as broken access control, business logic errors and gaps in sign in and session handling. Teams pick up to 20 findings, choose Fix with Agent, and get each fix back as its own pull request. On GitHub Actions the agent reads failed builds and retries up to three times, but it never merges anything itself.

Security teams no longer have to wait for the one engineer who knows the access model to fix every logic flaw, and reviewers still decide what ships. It needs a Hunter Agent subscription on the Enterprise plan, and model usage runs on the customer's own OpenAI or Anthropic key.

MCP / tool calling / API

Sourcegraph agents, such as Deep Search and the agent in Agentic Batch Changes, now support MCP through connectors. The same release redesigns the agent timeline so one card lists the questions in a thread, which helps on long Deep Search and Batch Change Agent threads.

Teams can plug MCP servers into Sourcegraph's code agents, so the agents can use outside tools while they answer questions or plan changes across many repositories.

Memory / state

Torq Retrospect moves a security team's past case records into Torq from a legacy SOAR, a ticketing system, another AI SOC or an MDR portal. An LLM works out the field mapping once, and the records are then copied as they are rather than rewritten, with unmapped fields counted and a sample checked by a person. Migrated cases arrive indexed with recent ones weighted higher, so Torq Recall can use them as precedent when triaging new alerts.

Switching to Torq no longer means an AI that knows nothing about the team's history, and the original records stay intact for audits. One customer moved more than 2,000 cases in under an hour.

Workflow orchestration

Inbound webhooks are now in public beta on Trigger.dev Cloud and the open source project, so events from Slack, Stripe, GitHub, Linear and more than 40 other services can start agents directly. Each webhook is verified, checked for duplicates and stored before it is routed to as many as 25 agents whose filters match. The dashboard shows every payload, filter, failure and replay alongside the steps of the agent it started.

Teams can build agents that react to the tools they already run without writing their own intake and retry layer, and a run waiting on an event or an approval sleeps for up to 90 days without compute charges.

Security / enterprise

Schedules and subscriptions in Agenta now run as the user or API key that created them, and pause when that identity can no longer run them. Evaluations run as whoever started them, and an approval keeps the identity of the run that asked for it while Agenta records who approved. In the same v0.124.0 release, removing a member, changing a role or deleting a key takes effect on the next request, and an API key can act only in its own project.

Security teams can tie every unattended agent run and approval to an accountable person or key, and taking access away from someone pauses their automations instead of leaving them running.

Workflow orchestration

Workflows 2.0 brings a rebuilt workflow builder to all Shapes customers, where a Workflows agent drafts a workflow from a plain description. Workflows can start on a new hire, a field change, a date such as a birthday or a manual run, with tasks, forms, signatures, emails, Slack messages and webhooks as steps. Every run can be followed in Activity, including from a person's profile.

HR teams can describe a people process and get a working automation to review, with run by run visibility and webhook steps that reach other systems.

Agent capability

Pi, the AI agent in Planhat OS, can now search the public web and draft and send emails from inside the conversation. A user can ask Pi to look up a company before a call, then have it write the follow up and send it. Admins decide whether each of these tools is available through Pi's tool controls.

Customer teams can go from account research to a sent email without leaving Planhat, and admins should settle the tool controls before Pi starts sending mail on users' behalf.

Agent capability

An xpander agent can now build new tools for itself: a SKILL.md skill, a connector generated from an OpenAPI spec with its OAuth2 sign in, or a desktop connector that records the steps of an app with no API. Everything an agent builds starts as a draft. Admins test and approve these drafts in the skills registry and decide who may build them.

Agents can close their own tool gaps, including apps with no API, without an engineering ticket, while nothing they build goes live until an admin approves it.

Agent capability

Version 1.1.21 of the alfred_ iPhone app brings the web Home to the phone, with routines shown as cards, swipeable cards for what needs the user and an Ask alfred_ bar. Two new Home Screen widgets come with it. One counts down to the next meeting with its attendees and a Join button and flags any reply owed to someone on the call, and the other shows the latest result of a chosen routine.

Client facing staff can see before a call that they still owe someone on it a reply, without opening the app, which is where follow ups tend to slip.

MCP / tool calling / API

API keys and MCP tokens in Searchable gain an Allow write access switch, so a connected AI assistant can make changes in a project as well as read its data. The switch is off by default, the key list shows which connections are read only and which can write, and write actions still follow the user's project role. Looker connections stay read only.

Teams can let outside agents act on their AI visibility work, not just report on it, while read only stays the default and project roles still cap what each connection may change.

Workflow orchestration

Administrative runs, new in Bernstein v3.23.0, give tasks that people or outside clients claim by hand an audited container, with no planner or worker agents started. Role sessions can now be told about new tasks as they are created through a live event stream. Tasks created directly over MCP or REST also join the signed claim and completion record that orchestrated runs already use.

Teams that hand out agent work manually or from their own tools keep the same signed audit trail as fully orchestrated runs, though the event stream does not replay missed events and is not a durable queue.

Browser/computer use

Apps built on Remy can now launch and control real browser sessions straight from their backend code. A new browser client opens a session and drives it step by step. On the same day, Remy workspaces gained the option to sign a Business Associate Agreement and run as HIPAA workspaces, where voice sessions, inbound calls and phone number purchases are refused.

Builders can ship apps that work through websites with no API without running their own browser infrastructure, and healthcare teams get a HIPAA mode that blocks the features the agreement does not cover.

Browser/computer use

Agentic RPA in Appian RPA 9.27 takes a goal written in plain language and works toward it on screen, capturing what it sees, then navigating, clicking, typing and reading until the goal is met. It works with any application visible on screen, from SAP and mainframe terminals to legacy desktop software and web browsers, and adapts when a dialog moves, a button label changes or a pop up appears. New Switch and Try-Catch actions can wrap it so a robotic task recovers from errors without stopping the run.

Teams with legacy desktop and terminal systems that have no API can automate them without configuring every step, and robots that adapt to screen changes need less upkeep than step by step scripts.

Workflow orchestration

Dify 1.18.0 adds Output Routes, so an Agent node can choose which workflow branch runs next based on a selection condition set for each route. Plain Agent v2 nodes now stream text as it is generated. Workflow Tools called from a workflow run as child graphs that pause, resume and stop together, and their steps can be opened from run history.

Teams can mix agent judgment with fixed workflow paths instead of choosing one or the other, and can trace what a nested workflow did when debugging an agent driven flow.

Integrations

Interview notes from GoodTime's AI Notetaker now flow into Greenhouse Scorecards. The notetaker takes word for word notes and sends the results straight back to Greenhouse for compliance and review, with no browser extension to install and no new scorecard templates to build.

Greenhouse customers get scorecards filled in from the interview transcript, which saves interviewers time and gives hiring reviews a more accurate record.

Workflow orchestration

Bretton agents can now be linked into one multi step investigation, where the outcome of each step decides which agent runs next. The next agent picks up the earlier step's documents, findings, narrative, outcome and analyst notes. Analysts start and follow these runs from a new Workflows page.

Compliance teams running a review such as enhanced due diligence, then a request for information, then a final decision no longer link separate cases by hand, and each agent starts with the earlier work in front of it.

Agent capability

A Vault in Settings now holds logins, cards and contact info that Bots fill in to sign in to sites or check out. Bots can also forward an email from their own inbox with its formatting, images and attachments intact. In the same release, Auto-review asks for approval less often when the user requested the action in their own words.

Bots can finish sign ins and checkouts without the user typing details each time, so it is worth deciding what goes into the Vault and which actions Auto-review lets through.

Agent capability

Hightouch's agentic marketing platform now sends lifecycle messages itself, so campaigns its agents build can run end to end without a separate engagement platform. Delivery covers email, SMS, push, in app messages and WhatsApp, in batches or in real time, with consent and compliance controls and domain management and warming. Customers who are not ready to switch can keep sending through the engagement platform they already use.

Lifecycle teams can drop a tool from the stack and keep data, AI agents and sending in one place, or move over at their own pace.

Deployment / data residency

OpenBox's runtime governance platform for AI agents can now be bought on AWS Marketplace and run entirely inside the customer's own AWS account and VPC. A signed Helm chart and an AWS CDK template set it up on Amazon EKS, and endpoints can be limited to an IP allowlist or kept inside the VPC. Each session's proof certificate is signed with a key in the customer's own AWS KMS or through an external attestation endpoint.

Banks, hospitals and public sector teams can put a policy check in front of every agent action while prompts, outputs and audit records stay in their own environment, and pay for it through their existing AWS billing.

Workflow orchestration

Coral 1.3.0 hands the deadlines for a Run, its approvals and its connections to timers kept by the Run Owner. A sweep now finds Runs whose lease has expired and recovers them, and Runs are drained when a server shuts down. OAuth sign ins can also finish on any Coral process, not only the one that started them.

Agent runs that are waiting on a person's approval or sign in are less likely to be lost when a server restarts, which matters once Coral runs long tasks for a whole team.

MCP / tool calling / API

Version 1.15.29 of CrewAI moves its agent executor out of experimental status and adds call: tool, which runs integration app actions directly. The same release lets the evaluation command run a crew and return a verdict in one step, exports evaluations as files, and adds an Azure DocumentDB vector search integration.

A stable executor lowers the risk of building production agents on CrewAI's core loop, and one step evaluation verdicts make it practical to gate a release on test results.

Workflow orchestration

Subagents that failed or were canceled can now be resumed in Goose v1.56.0 instead of started over. The UI links to each subagent's session, which opens as view only, so users can see what that subagent did. A new setting caps how many subagents run at the same time.

Long multi agent runs survive a single failure, and teams can keep parallel work within their model rate limits and budgets while still inspecting every delegated step.

Integrations

Hebbia now connects to Databricks, so a firm's own data such as portfolio metrics, positions and fund performance can sit in the same question as its documents, filings and licensed market data. Hebbia reads the data where it lives without copying it out. Each user signs in with their own Databricks account and sees only what they already have access to, and every answer shows its source.

Deal teams can bring warehouse data into diligence, for example checking a company's latest KPIs against its covenant thresholds, without filing a request with the data team. It is available to Hebbia customers with a Databricks account.

Deployment / data residency

FIPS mode is now available in self hosted Government deployments of Langfuse from v4.71.0. It runs Langfuse's cryptography, including TLS to its databases, hashing and application encryption, on FIPS 140-3 validated modules, and the app refuses to start if the host is not in FIPS mode. From the same version, the web and worker images for every deployment are built on Iron Bank's hardened Red Hat UBI 9.

Government and FedRAMP teams that must use validated cryptography can now run Langfuse for agent tracing and evaluation inside their own boundary, with module certificate numbers to hand to an assessor.

Deployment / data residency

WorkHQ SaaS services are now available in two new regions. Canada Central runs in Canadian data centers so customers can keep their data in the country, and a FedRAMP authorized GovCloud region serves U.S. federal agencies and federal contractors.

Canadian organizations with data sovereignty rules and U.S. federal buyers can now run WorkHQ automation in a region that meets their requirements. Blue Prism Assistant is not yet available in the Canada region.

Human approval / guardrails

Approvals in Serval no longer need a workflow behind them. Whoever is working a ticket can open an approval card on it, name who the approval is for, choose one or more approvers, add context and decide whether a justification is required. Approvers are notified as with any other Serval approval, and the decision is logged on the ticket thread.

IT teams can put a sign off on one time actions, such as letting a departing employee keep a wiped laptop, without building a workflow first, and the record stays with the ticket for audit. Admins turn it on under Capabilities in Settings.

Human approval / guardrails

Staff can now reply inside a conversation that an AgentFlow AI agent is handling, and the agent moves to standby instead of leaving. It keeps reading the thread but sends nothing while the person has control, and the person can hand the conversation back with a note. Every handover is recorded in the conversation, and a new AI handling inbox lists every conversation an agent is in.

Support teams can step into a live AI conversation and pass back facts the agent could not see, such as a discount agreed on a phone call, without the agent losing the thread.

Agent capability

CompanyOS, a new part of AgentOS, gives agents ongoing areas of responsibility called Products, each with Sources such as repositories, documentation, local Markdown files and websites. Scheduled Routines read those Sources, check existing work and propose Projects and Tasks, and agents carry out the work once a person approves it. Company Manager and OS Manager agents help set up Products and Routines and build the agents that do the work.

Teams can hand agents standing jobs like keeping docs current or turning customer feedback into fixes, without writing a prompt each time, while people still approve what gets done.

Observability / auditability

Classifiers in Confident AI can now listen to voice agent audio directly, catching interruptions, awkward silences and tone problems that a transcript never shows. A review step lets a person play the clip and confirm a label before it sticks. The rebuilt audio player marks each issue on the timeline, and Pipecat pipelines can now send their audio into traces.

Teams running voice agents can find a bad moment in a production call and replay it, instead of guessing from the transcript, with a human check on what gets flagged.

MCP / tool calling / API

MCP servers built with Mastra can now send events to connected clients by push, poll or webhook, instead of only answering requests. Agents receive each event in their thread as a notification, and workflows can start a run from one, so a failed deploy or a merged pull request can set work in motion. The feature is experimental, needs @mastra/mcp 2.5.0 or later, and records event traces automatically when observability is set up.

Teams can build agents that react to changes in outside systems without running a separate pub/sub service or writing their own polling code.

Integrations

A new Netlify connector for Grok lets a Grok agent answer questions about Netlify projects, update supported project settings and deploy through conversation. It works across Grok Bot, Grok CLI and grok.com, and Grok calls it on its own when a task needs Netlify. Users add it from Connectors in Grok.

Teams that build with Grok can ship and check deploys without leaving the agent, which puts Grok alongside the other coding agents that can deploy to Netlify.

MCP / tool calling / API

Codex CLI 0.164.0 lets enterprise MCP connections authenticate with eligible Codex account grants or a dedicated sign in, including OAuth clients that require secrets. The same release masks credentials during supported remote Bash commands and Zsh snapshot sessions, including credentials from custom providers.

Security teams get a managed way to connect Codex to enterprise MCP servers, with less chance of secrets showing up in agent command output on remote hosts.

Observability / auditability

Event Telemetry records each step a Giga agent takes, including calls, actions, code runs and API requests, as ordered events that teams pull as JSON with the Giga CLI. Events carry the agent version that handled a conversation, say why a transfer failed, and can be read back for up to 30 days. Scout, the assistant in the Giga console, reads the same events to explain what went wrong in a conversation.

Support teams can tell whether a failed request broke in the agent or in their own order or phone systems, so the ticket goes to the team that owns the fix instead of landing on the agent.

Integrations

Recruiters and hiring managers can now check pipelines, prepare for debriefs and draft scorecards inside Claude with Greenhouse in Claude, a plugin built on Greenhouse MCP. Beyond answering questions from live hiring data, it can update candidates, move them forward, schedule interviews and create jobs and job posts. It is in open beta for Greenhouse Recruiting customers on Core, Plus and Pro plans, and works in Claude, Claude Cowork and Claude Code.

Hiring teams can do routine work from the assistant they already use, while each person signs in under their own Greenhouse permissions and admins decide which information and actions are available.

MCP / tool calling / API

Version 1.1.0 of the Hyperspell MCP server adds index_website, a tool that lets an agent start a bounded, one time crawl of a public website. A companion tool reports the latest crawl for the caller, including partial failures and the page limit.

Agents can bring a public website into Hyperspell on request instead of waiting on a separate ingestion job. On the hosted server the tools stay where crawling is not available for the caller's region or credential.

Observability / auditability

Langflow 1.13 adds opt in audit logging, an append only history of flow and project operations. Each event records who attempted a change, on which resource, when, how it completed and what was written. Once switched on, events can be queried through separate API endpoints for flows and for projects.

Teams running Langflow for several users or under regulation get a record of who changed which flow, a common audit requirement.

Agent capability

OpsAI, the SRE agent in Middleware, now comes with built in skills: reusable instructions for investigations such as error analysis, latency troubleshooting, Kubernetes, databases and LLM observability. Users pick a skill by typing / in OpsAI chat, import or write their own, and the agent also chooses relevant skills on its own. A new project context setting holds notes on architecture, infrastructure and requirements that OpsAI applies to every new chat and issue investigation.

SRE teams can write their own runbooks and environment facts into the agent once, so its investigations and suggested fixes follow how their systems actually work rather than generic advice.

Workflow orchestration

Automation Builder lets regulatory teams set up workflows that start from a regulatory alert or on a schedule. AI agents in those workflows assess impact, assign follow up tasks, flag active Label Compliance projects, send stakeholder communications and carry out later steps, and each run records its outcome. Regulatory decisions stay subject to human review.

Teams can go from a new requirement to assigned work without manual hand offs, and the record of each run gives reviewers a trail of how every alert was handled.

MCP / tool calling / API

The hosted Causely MCP server now carries its own guidance on how to use its tools, so a new background agent works well from its first connection. Every agent action can be traced back to the diagnosis and evidence it was based on, and agents start from structured results on what is wrong and what recently changed instead of scanning the environment. The server is now listed in the GitHub MCP Registry, with example background agents and MCP Skills published as open source.

SRE teams get a pull request or Slack message from an agent with a trail a person can check, and spend less prompt work and fewer tokens getting each agent going.

Agent capability

Agents built in Creatio AI Studio can now work in the Creatio Feed of a record, a channel or a single post through a new Creatio Feed skill. An agent reads the feed for context and, when the user asks, posts, comments, edits the user's own posts and mentions colleagues. When someone asks AI Twin to build an agent that works with the feed, it adds the feed instructions on its own.

Sales and service teams can have agents take part in the conversations around a deal or case, not only read and update the record itself.

Agent capability

Vulnerability watch, new in Roomote 1.19.0, checks connected repositories against official CVE records and starts remediation assessments that propose fix pull requests for review. It is opt in and off by default, and admins can exclude repositories and see findings, coverage gaps and the assessment backlog. Fixes are never merged automatically, and dependency release age policies still apply.

Security and platform teams get an agent that does the legwork on known vulnerabilities while a person keeps the merge decision.

Workflow orchestration

Agent chats in Taskade now keep working after the user closes the tab or reloads, and the agent finishes its reply. Reopening the chat picks the reply up live where it left off, a second tab can follow the same reply as it streams, and Stop halts the agent from any tab.

Teams can start a long research or build task with an agent and walk away without losing the work when the browser closes.

Agent capability

AgentSpot is a new ThoughtSpot workspace where anyone describes a job in plain language and gets an agent, a workflow that runs on a schedule or trigger, or an interactive app, all built on ThoughtSpot's semantic layer. Agents act within each user's permissions, every action they take is logged, and memory is kept at company, agent, user and agent user levels. A free plan includes unlimited users and agents, and paid usage is charged per conversation, workflow run or app session rather than per token or seat.

Business teams can build an agent once and share it across the company on agreed metric definitions, while IT keeps control through permissions and the audit trail and finance avoids token based bills.

Human approval / guardrails

Release 5.9 lets teams decide, tool by tool, whether a Blits.ai agent runs it on its own, asks the customer to confirm, or waits for an employee to approve. Approvers get an inbox and a review button on live conversations, and agentic workflows pause while an approval is pending, then resume once it is decided. Write actions on business system integrations now need approval by default.

Regulated teams can choose exactly which agent actions need a human sign off before anything is written to a core system, and new integrations start from the safe setting.

Agent capability

Pigment Agents can now search and read every file stored in Pigment Files, instead of being sent each one by hand. Company wikis, rules, style guides and policies become available to Agents at once, so an Agent can summarize a policy PDF or pull figures from a contract while a planner builds or reviews a model. A Security Admin turns the access on in security settings.

Planning teams get agents that work from the company's own policies and contracts without repeated uploads, and security keeps a single switch over whether that access exists.

Browser/computer use

Agent Browser gives an AI agent on Telnyx its own isolated Chromium session through an API, so it can use websites that have no API. A session lasts 15 minutes by default, up to 30, and can only reach a list of up to 50 allowed domains set when it starts. The agent reads each page as labeled elements and acts by reference with ten action types, including navigate, click, fill and screenshot.

Agents can book slots or fill vendor forms on sites with no API, and security teams get the reachable sites and session lifetime fixed before the agent begins.

Browser/computer use

Viktor's Screen is now open to every workspace: users can watch what Viktor is doing in the browser, take over, and hand control back. Saved logins stay private to the person who saved them unless they choose to share one with the whole workspace. The same update adds Airwallex to the integration catalog through its AgentOS MCP, so Viktor can work with payments and accounting.

Any team can now supervise and step into the agent's browser work on sites with no API, while keeping control over whose credentials it uses.

MCP / tool calling / API

The Government & Legal Index is a free search index of 30 million pages from US federal, state and local government sites, covering statutes, regulations, city codes, permits, court opinions and agency guidance. Agents reach it through a new government search endpoint in the API, a matching tool in the Firecrawl MCP server, the CLI, or the Firecrawl plugin in ChatGPT and Claude. Searches use no credits.

Agents answering legal, tax or compliance questions can cite primary sources instead of secondhand summaries, and every result links back so the agent can check who issued it and which version it is.

Workflow orchestration

Forest Workflows can now start on their own when a record changes, such as a status update or a received document. Service accounts give the tools and AI agents that act in Forest their own role, teams and revocable application tokens, and inbox automation lets AI agents work inboxes under that identity. Runs that need a person or fail go to a fallback inbox, where an operator picks up from the step where the run stopped.

Regulated ops teams can let agents and outside systems run processes without borrowing a person's login, while every automated run stays traceable to the account it acted as and the channel that started it.

Agent capability

OCRia-1, a 4 billion parameter vision language model that LlamaIndex trained for document parsing, now powers LlamaParse at the same price. It rebuilds complex tables with nested headers and merged cells, turns charts into tables of their data points, and keeps formatting such as superscripts, checkboxes and math notation. On ParseBench it scores 87.4 overall, 12.1 points ahead of the next parsing model under 35 billion parameters.

Agents that read financial statements, insurance schedules or forms get more accurate input from LlamaParse with no migration or price change, which matters most where a misread table becomes a wrong number downstream.

MCP / tool calling / API

Conduit 0.16.0 adds three tools that every MCP client always sees, for listing, describing and running whatever tools the user can reach right now. Clients that never refresh their tool list, such as Codex or Claude Code's deferred tool search, can now find and use a tool an admin enables mid conversation. Calls made this way go through the same policy, sign in flow, usage tracking and audit as a direct call.

Newly approved connectors reach users without restarting their agent sessions. A client's per tool approval rules now see only the generic runner, so approving it once approves whatever it runs next, and admins should review those rules.

MCP / tool calling / API

Outside agents connected to Zite, such as Claude, ChatGPT or Cursor, can now upload files from the user's computer straight into the database and create or edit AI fields. Their changes appear in a feed of their own, and version history shows which agent made each one. A new Bring your own agent link on Home connects an agent without opening a workspace first.

Teams that run Zite from their own chat or coding agent can hand it more of the work and still see afterward exactly which agent changed what.

Human approval / guardrails

Agents running in Browserbase can now sign in to websites through 1Password Agentic Autofill, added through the Browserbase extension system. When an agent reaches a sign in form it asks for a specific login, the user approves or denies the request in 1Password on their own device, and 1Password fills and submits the credentials inside the browser. Filled values are removed after submission.

Agents can work in dashboards and SaaS tools behind a login while passwords never pass through the application, the model or the logs, and a person signs off on every credential request.

Agent capability

A new version of CoCounsel Tax, Audit & Accounting, now available to customers in the United States, is an agentic platform that plans the work and picks the tools for it. It writes and runs code for calculations such as depreciation and qualified business income, builds multi tab Excel workbooks with working formulas, and edits the spreadsheets and documents a team uploads. Native MCP support lets it pull context from other systems, alongside live integrations with UltraTax, Engagement Manager and Guided Assurance.

Tax and audit teams move from an assistant that answers questions to one that hands back finished workpapers they can review, with exact calculations instead of estimates.

MCP / tool calling / API

Mantle agents can now pass context to MCP servers as metadata kept apart from the tool arguments and the model's history. Values held in skill or project memory can be sent with every tool call, and a hook can attach extra metadata to a single call. When a tool needs confirmation, the metadata is saved and sent again once the call resumes.

Builders can hand customer, tenant or session details to their MCP tools without exposing them to the model. It ships in the 3.22.0.dev1 preview build of Mantle, which is itself in open beta, and the metadata stays in conversation history, so it should never carry secrets.

Integrations

Findings from Cyera's data security posture management platform now flow into 7AI, where its investigation agents work each exposure knowing what data is at risk and rank it by business impact rather than severity alone. The link is a read only API connection, with nothing to install on the Cyera side.

SOC teams that run Cyera can hand exposure, risky access and AI usage findings to agents that already know which data is regulated, instead of gathering that context from each tool by hand.

MCP / tool calling / API

A per step model router in the Blackbox AI gateway sends each step of a coding agent session to a closed frontier model, a big open weight model or a small open weight model, depending on how hard that step is. Microsoft Decision-1 judges each step in about half a second, and Blackbox engineers configure the routing for each agent workload. In live Terminal-Bench 4.0 runs the router completed 65 percent of tasks at 38 percent lower cost per task than sending every step to Claude Opus.

Most agent steps, such as listing files or rerunning a test, do not need a frontier model, so teams running coding agents through Blackbox can cut spend on long sessions while keeping top models for the hard steps.

Deployment / data residency

Version 10.140.0 of the on premises release adds Performance versions of some F5 AI Guardrails, starting with PII core and the Adversarial attacks package, the new name for the Prompt Injection package. Performance guardrails use smaller ML based scanners that run on CPU systems, or on GPUs for higher throughput. Comprehensive versions keep the GenAI based scanners, which apply deeper contextual reasoning and still need a GPU.

Self hosted teams without spare GPU capacity can now run core guardrail scanning on CPU hardware, and choose per guardrail between faster scanning and deeper detection.

Agent capability

Stakeholder search arrives on an account's People page, where users ask Mutiny to find the people involved and then accept or decline each suggestion. Mutiny shows where each suggested person's details came from and asks before it runs another paid search.

Reps can map the people on an account without leaving the workspace where they build assets for it, while a person decides who is added and approves any further data spend.

Agent capability

Agent Studio in RunMyJobs 2026.4 is a guided builder for reusable AI agents that run as steps inside workflows. Builders pick a model, write instructions and grant access to specific tools from connected MCP servers with token and runtime limits, then test the agent under production controls while write actions are simulated. Published agents resolve recognized exceptions within the permissions granted and escalate unfamiliar cases for human review with the investigation attached.

Operations teams get agents that live inside governed job chains with explicit tool rights, spend limits and a human escalation path, under the security partitions, role based access and audit model RunMyJobs already applies.

Integrations

A personal WhatsApp account can now be linked to Strawberry by scanning a QR code in Settings > Integrations. The companion then finds chats, searches messages, and sends replies, photos, documents and voice notes, with sending governed by the user's app permissions so approval can be required first. The same update lets routines that start from an app event carry an Only if condition, so only matching events start a run.

People who work with customers over WhatsApp can hand that channel to the companion and still decide which messages need their sign off before they go out.

MCP / tool calling / API

Turbine 26.5.0 lets the Hero AI Companion and Hero AI Native Actions call tools from external MCP servers that an administrator registers, in both Turbine Cloud and on premises deployments. Turbine can also run as an MCP server of its own, so AI clients such as Claude, ChatGPT, Cursor and Gemini CLI can call the components an administrator chooses to publish. Role based access control still applies to those components.

Security teams can bring outside tools into AI investigations and let the AI assistants they already use run selected Turbine components under existing access rules, without custom connectors. OAuth is not supported in this release.

Agent capability

Tabnine 6.7.0 opens the new Tabnine Agent to every team with Agent access, and IDE users can switch to it from the introductory screen. The agent gains a read only Chat mode with one click Apply, image and file attachments, and built in read only Jira access. The same release lets administrators set an organization wide monthly Agent budget and hide usage costs from end users.

Any team with Agent access can move developers to the new agent now, and the read only Chat mode plus a monthly budget cap give admins a lower risk way to widen use.

Agent capability

Context compaction and session notices now work for external agents that connect to Zed over the Agent Client Protocol (ACP), starting with release 1.25.2. A third party agent running in the editor can compact a long conversation and show its session notices inside Zed.

Teams that bring their own coding agent into Zed through ACP can keep a long session going in one thread instead of restarting it as the conversation grows.

Workflow orchestration

Any Anvilogic Blueprint, the layer that chains AI agents into a repeatable security operations workflow, can now run on a schedule instead of waiting for someone to click Run. The owner sets a cron schedule, a time zone and a maximum runtime, and runs go ahead whether or not anyone is online. Each scheduled run uses the owner's permissions, pauses at approval steps until a person signs off, and is skipped if the previous run is still going.

Weekly detection tuning, daily hunts and cost reviews happen on time without an analyst remembering to start them, and every run shows up in history with its status, duration and credits used.

Integrations

Version 3.59.0 of the Genius Chat Widget extends its Add to cart button beyond Shopify, so shoppers can add a recommended product to their cart without leaving the chat. Stores on Salesforce Commerce Cloud or a custom storefront describe their own cart request in the widget's server side config, or hand the job to a callback on their page.

Retailers outside Shopify can let the shopping agent finish the sale in the conversation instead of sending the shopper off to a product page.

Memory / state

Latenode Tables, a table database built into the platform, is now open to every user, with a free plan already linked to each account. Scenarios reach it through a native app that can create, search, update and delete rows, including in batches, and upload files. A New Row Created trigger starts a scenario whenever a row is added.

Teams building automations and agents on Latenode can keep records and working state next to their scenarios instead of wiring up an outside spreadsheet or database.

Agent capability

Tax Center in the Docyt Accounting Suite builds a separate set of tax basis books for each year from the ledger a business already closes, and checks that they tie to the operating books first. For cash method returns it suggests where each balance sheet account belongs, flags anything it cannot place, and drafts the accrual to cash conversion entries for review. It exports a return ready package with a book to tax reconciliation, and gives the outside CPA a Tax Accountant role with access to the tax books only.

Finance teams stop rebuilding the cash conversion in a spreadsheet every year, and a filed year locks with a record of who carried it in and who filed it. Docyt prepares the package but does not file the return.

Workflow orchestration

The Prompt Engine in Gnani.ai's Agent Builder splits a voice agent's conversation into scenarios, each a self contained piece of the flow, joined by transitions set up in advance. On every turn the agent reads only the scenario it is in, not the whole script. Global transitions catch moments such as a callback request, a competitor mention or a request for a human, wherever the call happens to be.

Long branching calls such as insurance renewals or claims can add cases without each turn getting slower or costlier, and builders can fix one part of a conversation without touching the rest.

Observability / auditability

Airtable's audit log now shows whether a person or an agent took each action, and in which base. Actions by Omni, Airtable's own assistant, are attributed directly, while actions from outside AI tools trace back to the integration and the person who connected it. Events inside a base, such as creating a table, updating an interface page or deleting a comment, are rolling out now, and the log can be filtered, exported, or streamed through the audit API into security tools.

Admins can answer who or what changed a base before they let agents work on production data. Audit logs remain an Enterprise Scale feature.

Human approval / guardrails

Composio Firewall checks every action an agent takes through Composio against the user's intent and the limits set by the developer and user, before the action runs. It also screens tool results before they reach the agent and blocks any that carry malicious instructions. It is live for all Composio developers.

Teams giving agents broad tool access get a prompt injection screen and an intent check at the point of execution without building their own, which closes part of the gap between what an agent is allowed to do and what the user asked for.

Deployment / data residency

Secure AI Agents, Idira's product for giving AI agents governed identities and access, now runs in three more AWS regions: Frankfurt, London and Singapore. Tenants in London and Singapore can also ask their Idira representative to turn on database access for agents through the Secure Infrastructure Access MCP server.

Organizations in Europe and Asia Pacific can keep agent identity and access controls in their own region, which removes a residency blocker for regulated buyers.

Agent capability

AI Investigation, now live for all users, puts an agent on each test failure while the suite is still running. It clears environment causes of flakes and re runs the test, repairs broken tests while keeping the same assertions, and reproduces real bugs in its own browser to file a report with video, traces and repro steps. One lead agent routes related failures to the same investigator, so one fix can clear several tests at once.

QA teams get failures triaged and often fixed during the run instead of after it. They can dispute any conclusion, follow along from ChatGPT or Claude through QA Wolf MCP, and steer investigators with skills kept in the workspace.

Agent capability

Eval Health is a new managed agent in Arize AX that reviews the evals on a tracing project. It reads recent spans, tasks and evaluators, then files an issue for each coverage gap or broken eval with a proposed fix: create, repair, delete or optimize an eval, such as turning an LLM judge into a code eval to cut cost. Nothing changes until the user accepts the proposal.

Teams running evals at scale get an agent that keeps eval coverage and cost in check, with a person approving every change before it runs.

MCP / tool calling / API

BYO (Custom MCP) lets an organization register its own MCP servers and connect any MCP compatible tool to Harvey. Workspace admins decide which tools are connected and who can use them.

Legal teams can bring internal and niche systems into Harvey without waiting for a connector from the vendor, and access stays in the hands of their own admins.

Observability / auditability

AI spend in Haystack Enterprise Platform can now be broken down by model, agent, service and team, both in central reporting and next to production traces. Teams can spot the expensive or unreliable steps in an agent and compare models and harness setups. Each change can then be tested for whether it lowers cost while still meeting the quality the task needs.

Teams running agents in production can see which agent or team drives the bill and test a cheaper model against a quality bar before switching.

Agent capability

Startups on Stripe Treasury can now let agents such as ChatGPT and Claude run their business finances through Stripe's MCP server. Agents can check balances, pay invoices, move money between accounts and convert currencies. The business chooses which accounts and data an agent can reach, whether it can only read or also act, and when money movement needs a person's approval.

Founders can hand routine treasury work to an agent and still approve each payment before money moves. Spending limits for agents are not available yet.

Pricing / packaging

SaaS and hybrid customers can now add curated open weight and frontier models that run on TrueFoundry's own capacity, with no provider account to create and no API key to rotate. Usage draws down a prepaid credit balance that does not expire, starting with free credit and no card, and an optional automatic top up. Calls still pass through the AI Gateway with tracing, cost tracking and access control.

Teams can get agents running on gateway models before any provider contract is signed. Credit is billed apart from the plan invoice and managed models stop serving at a zero balance, so someone needs to own the top up.

Agent capability

Version 2026.10.19 of the Equal iPhone app adds Make a Call, so the assistant now places calls for the user instead of only answering and screening incoming ones. The same release adds Ask Equal for talking or chatting with the assistant anywhere in the app, and Instruction Packs, ready made sets of call handling instructions applied in one tap.

Equal moves from screening calls to starting them on the user's behalf, which widens what can be delegated to it and makes its instruction settings matter more.

Agent capability

AI Employees from Kognitos are agents assigned to 61 finance and banking roles such as AP Specialist, Reconciliation Accountant and KYC & Onboarding Analyst, backed by a library of 515 standard operating procedures. Each role has defined limits on what it may do without human approval and never adds a step that is not in its approved SOP. A builder agent, Quill, drafts new SOPs in plain English from a team's systems and recent transactions, and nothing is published until a person approves it.

Finance teams get agents whose steps are fixed by an approved procedure, with a run journal an auditor can read, which answers the audit objection to agents that choose steps as they go. AI Employees are in preview for Kognitos customers, so confirm which roles are ready for production use.

Agent capability

ReadySlate from Findem uses AI recruiting agents to turn job requirements into a slate of candidates who are interested, qualified, verified and ready for on site interviews. In ReadySlate Enterprise, candidates the customer unlocks move into its ATS, VMS or HRIS, including Workday, Greenhouse and iCIMS, and the customer pays only for those it unlocks. ReadySlate Lite offers a self service version in Findem Studio for startups, small businesses and hiring managers who run their own searches.

Hiring teams can hand the run from requirements to a screened slate to agents while flagged candidates still go to a person for review, and Enterprise cost follows unlocked candidates rather than seats.

Integrations

Noma connected its AI-DR runtime product to NVIDIA OpenShell through the supervisor middleware, so each request an agent makes inside an OpenShell sandbox, including prompts, tool calls and MCP traffic, passes to Noma for inspection after OpenShell's own policy check. When a Noma policy triggers it can block the action, steer the agent, mask sensitive data or alert a person, and the verdict is enforced outside the agent's process.

Teams running Claude Code, Codex or their own agents in OpenShell sandboxes get content level checks on traffic that policy already allows, such as leaked secrets or injected instructions, without changing the agent.

Integrations

The Meta Ads integration in Viktor can now create new lead forms or copy existing ones, and can list, inspect and create value rule sets that adjust bids by audience, placement or location before attaching them to rebuilt campaigns. Every new form and every new value rule set waits for the user's approval, and a copy refuses to go ahead if any part of the source form cannot be copied whole.

Paid social teams can hand form swaps and bid adjustment setup across many ads to the agent instead of clicking through Ads Manager, while anything that affects spend still stops at a human approval step.

Integrations

Command Zero went live as a launch partner for Tenable OPEN for AI with an Exposure to Verdict investigation workflow. A priority exposure finding from Tenable feeds into a Command Zero investigation that checks whether the exposure was actually exploited, with the Tenable Hexa AI MCP Server and Command Zero's MCP server working in the same AI session. The approved verdict goes back to Tenable, and the workflow is published as a documented workflow and a ready to use skill on Tenable's CyberAgents Exchange.

Exposure and SOC teams can learn whether a finding was exploited without switching consoles, and every step is recorded and reviewable before the verdict is written back.

Agent capability

Duo Accounts is a new account research agent in Amplemarket's Duo suite. It combines CRM activity, Duo signals and Amplemarket company data with public sources to give each tracked account reasons to engage, each linked to its source, plus a relevancy level of high, medium, low or none, refreshed on a schedule. A Plan outreach step hands the research to the user's AI assistant to recommend stakeholders and angles, which the rep reviews before acting in Amplemarket through MCP.

Account teams get a ranked list with the evidence shown, so a rep can check a reason before acting on it. Enablement goes through the customer's CSM or account manager, so plan eligibility is worth confirming.

Integrations

The Salesforce integration now syncs both ways for Leads and Contacts, including updates and custom variables, and logs outreach on the matching Salesforce record. Activity sync covers sequence emails, replies, LinkedIn messages and connection requests, bringing the integration level with Salesforge's HubSpot connection.

Outreach that Salesforge's agents and sequences send, LinkedIn activity included, now lands on the CRM record without manual logging, which matters for teams whose reporting and handoffs live in Salesforce.

Agent capability

Medical benefit drug utilization management now runs on Cohere Health's Cohere Unify platform, so drug and medical services prior authorization now share the same clinical intelligence layer. Its AI assembles the documentation and plan specific criteria for each drug review, automates policy aligned requests in real time and routes complex cases to expert reviewers with the clinical context attached. The capability is available now to new and existing health plan clients.

Health plans can run medical drug and medical services prior authorization on one platform instead of adding a separate drug utilization management system. Support for the CMS medical drug authorization requirements is promised once that rule is final, so it is not available yet.

Agent capability

Scaled Cognition Voice, a text to speech model built for live customer calls, is generally available and streams audio through a standard API to the company's APT agents, third party agents, contact center platforms or legacy IVRs. The company states 70 ms latency and more than 25 languages, and voices can be cloned from a short recording. It runs self hosted in the customer's own cloud or data center at $1 per million characters, or hosted by Scaled Cognition at $9 per million characters.

Voice agent buyers in regulated sectors can keep audio inside their own environment and adopt the voice without moving their agents to APT. Latency figures are vendor statements without a published test method.

Agent capability

Maia, the AI that builds scenarios in Make, now has a Plan mode that asks its questions up front, lays out the modules, assumptions and approach, and builds nothing until the user approves, checking connections, app capabilities and plan tier access along the way. Maia also gains Memory, which carries a user's preferences between scenarios and can be edited or erased, image attachments for building from a sketch or screenshot, and Resolve error, which reads a failed run's data and proposes a fix the user reviews before saving.

Teams get a reviewable plan before an AI builder changes a scenario and a direct path from a failed run to a proposed fix, which lowers the risk of letting an agent edit production automations.

Security / enterprise

MCP Management, now generally available from Celigo, puts MCP servers built in Celigo and those from third party vendors behind one MCP Gateway. Access is set tool by tool and tied to the customer's identity provider, so every call maps to a verified user rather than a shared credential. An MCP Registry lists every MCP server in the organization and which assistants and agents use it, and lets employees request tools that administrators approve, with each approval recorded.

Security and IT teams get one place to approve and audit what AI assistants and agents can do in business systems, often the step that holds AI pilots up in security review. Because Celigo runs the tools rather than only routing requests, teams can see which systems an agent touched and what it changed.

Human approval / guardrails

Approval and input requests that pause an Elastic workflow can now reach people through Microsoft Teams and ServiceNow comments. Delivery through Slack arrived on October 8 and through email and HTTP on October 15, so an approver can answer in the tool they already use and the workflow resumes.

Approval gates on agentic automation only hold if approvers see the request in time, and Teams and ServiceNow reach the people who sign off on security and IT changes without asking them to watch Kibana.

Agent capability

Kit, a new AI assistant inside ChurnZero, answers questions about one customer or a whole book of business from account data, internal documentation and the ChurnZero Knowledge Base. Users can put the answer to work by creating tasks, drafting personalized outreach or saving account lists as dynamic segments that launch a play or journey. It sits in the top bar on every ChurnZero screen and in Slack and Microsoft Teams.

Kit works under existing permission groups, admins choose which knowledge it uses, and nothing reaches a customer until a team member reviews it, so customer success teams get an assistant that acts without copying customer data into general purpose chat tools.

Integrations

A Datasite connector in Hebbia lets deal teams sign in with their own Datasite credentials and sync one or more virtual data room projects, bringing in only the documents each person has permission to see. Teams can then run a full diligence question list across a data room and turn the findings into a diligence summary or IC memo, with every point cited to the source page.

M&A and capital markets teams can analyze confidential data room files without exporting them, with Datasite permissions carried through per user. It is available to Hebbia customers with active Datasite projects.

Agent capability

Speech to Text from Rime is a new streaming and whole file transcription service for voice agents in English and Spanish, built from the same hand transcribed conversation data behind its voices. Rime reports a median of 105 ms from end of speech to final transcript and word error rates on the FLEURS test of 4.5% in English and 5.3% in Spanish. It is free for 30 days, until November 19.

Teams already using Rime voices can now get the listening and speaking halves of a voice agent from one vendor. The latency and accuracy figures are Rime's own measurements.

MCP / tool calling / API

The SnapLogic MCP Server now supports OAuth 2.0 browser login, so MCP clients authenticate with SnapLogic credentials instead of credentials stored in the client configuration, and refresh tokens keep long running agent sessions going. New tools let agents create project spaces and projects and delete pipelines, and execute_pipeline and validate_pipeline now return a conflict error instead of overwriting an existing pipeline with the same name.

Browser login takes stored credentials out of agent client configs and, per the release notes, enables connection through Anthropic's Claude Connectors Directory, while the overwrite protection lowers the risk of an agent replacing a production pipeline by name.

Human approval / guardrails

In Rasa Pro 3.21.0 the copilot can plan before it builds: a prompt produces a build plan listing the agent's skills, which the user can read, refine in conversation or edit directly before approving. Only approval scaffolds and builds the agent, and plans are versioned so every refinement is kept.

Builders get a review and approval checkpoint before an AI assistant generates agent code, with every plan revision kept as a record of what was approved.

Agent capability

SageOx now watches a team's Discussions for an open question that several coworkers keep returning to and, on its own, invites one of them to host a short whiteboard jam. The invited host can accept or pass, the others are asked only after someone says yes, and Recent jams shows every jam with its outcome. It ships as a Feature Preview turned on for every team, and an owner or admin can switch it off in Settings.

SageOx now starts meetings on its own initiative rather than only recording them, gated by host acceptance and a team wide off switch, so buyers should check that this proactive behavior fits their team norms.

Human approval / guardrails

Release 26.24 adds secret detection to Fiddler's Sensitive Information guardrail, flagging API keys, tokens, credentials and high entropy strings across more than 35 known secret formats as a new entity type in the PII categories. A validation layer also checks URLs, IP addresses, emails, SSNs and credit card numbers against format rules to cut false positives.

Agents that call tools and read files can leak credentials into prompts and responses, and Fiddler guardrail users now catch those secrets in the same check that handles personal data.

Workflow orchestration

Voice On-Call is now generally available to all Thread Voice AI customers. A Flow on any ticket condition triggers it to call down an ordered list of technicians until one presses a key to accept, then it adds that technician to the ticket, sets a chosen status and writes every call and acceptance as a ticket note.

MSPs can replace a separate after hours paging tool with a Flow triggered voice agent and keep a record on each ticket of who was called and when. It needs an Outbound Agent set up.

Integrations

Customers of both Thomson Reuters and Otto Schmidt can now use German law commentary inside the classic version of CoCounsel Legal through the Otto Schmidt Legal Data Hub. More than 120 modules of expert commentary sit in the international research workflow alongside Westlaw and Practical Law content, with citations, while support in the newer agentic CoCounsel Legal is planned for 2027.

Firms with matters that touch Germany can research German commentary in the same cited workflow instead of a separate publisher platform, given an existing Otto Schmidt entitlement.

Deployment / data residency

FlowX.AI 5.9.6, a maintenance release on the 5.9 long term support line, ships the document parser with its layout, table and OCR models built into the image. Nothing downloads at startup or on the first parse, so the parser works without internet access and on a read only filesystem.

Banks and insurers running FlowX in isolated networks can run document extraction for their agents without opening outbound access to model hosts.

Integrations

ADK 2.19.0 for watsonx Orchestrate adds a provider option to the agents discover command for Google Vertex AI, AWS Bedrock and Microsoft Azure AI Foundry. Using credentials stored in a connection, it lists the agents in those accounts and registers them as external agents in Orchestrate, with an option to limit discovery to named agents.

Teams with agents already built on the three major cloud agent platforms can bring them under one Orchestrate control plane without writing each agent definition by hand.

Agent capability

Ramp for Nonprofits includes Grant Policy Agent, which checks proposed expenses and bills against a grant's spending policy and flags likely mismatches before approval or allocation, such as a hotel bill against a grant that excludes travel. The launch also adds Grants Hub, grant budgets with spent, committed and remaining balances, program cards tied to a grant, and sales tax reclamation.

Checking spend against grant terms is manual work that small finance teams often skip, and an agent that screens each request before a person approves it lowers the risk of misallocating restricted funds.

Observability / auditability

The rebuilt Reports view in Neople Studio gives every answer the AI support colleague writes a single outcome, with weekly and monthly views, a custom date range and Count, Share and Trend modes. A Tickets tab tracks full automation, first reply automated, handed over and closed by the Neople, and every number clicks through to the tickets behind it.

Support leaders can see what the agent actually did with each answer and which topics still need people, without exporting data, which helps decide where to extend automation next.

Agent capability

Release M29 adds Agentic One Click Remediation, where the agent picks a predefined remediation playbook from a case recommendation and fills its inputs from the case and source alert for an analyst to launch with one click. Teams can also set auto remediation conditions so eligible actions run on their own, and those runs are listed in a separate Auto Remediated section.

The managed SOC agent moves from recommending containment to carrying it out under conditions the customer sets, so those rules become the control point to review. Playbook access still follows role based permissions.

Agent capability

Once Annie, Anyshift's SRE agent, has joined a Slack thread, she now decides for herself whether each new message is meant for her and replies without needing a fresh tag. Teammate chatter and ambiguous messages get no reply, and threads where someone pressed Stop on her last answer still require a tag.

Incident channels can keep a natural back and forth with the agent, and admins keep a single switch, on by default, to turn that judgment off.

Agent capability

Storage troubleshooting in Spark, Nexthink's employee facing IT agent, was scheduled to switch on by default for all tenants on October 19, after the 2026.8 rollout completed in every region. With it on, Spark looks at the files and folders on an employee's device to find what is filling the disk, can clear content taking up space, and suggests moving large files to OneDrive.

More disk space tickets can close without a technician. Tenants that never opted in now have Spark reading file and folder names on devices unless an admin turns that setting off, which privacy reviews should check.

Security / enterprise

Portal26 launched the Managed Intelligence Platform, a version of its AI governance product built for managed service providers. It lets an MSP discover the AI already in use across every client, manage the security and compliance risk that creates, and turn the findings into billable services such as license optimization and cost management.

Small and midsize businesses that rely on an MSP can get shadow AI discovery and governance as a managed service, rather than running an enterprise deployment themselves.

Agent capability

Three new tools in Quo's MCP server, start-agent-call, get-agent-call and cancel-agent-call, let an AI client have Sona, Quo's AI agent, place a real phone call from a Quo number to get something done, such as booking an appointment or confirming an order. The client confirms the calling number, the person to call and the task with the user before Sona dials, and get-agent-call returns Sona's summary once the call ends. The tools are in beta, and each call uses 50 automation credits, with calls under 15 seconds free.

Sona moves from answering inbound calls to placing outbound ones on instruction from any MCP client, so scheduling and confirmation calls can be delegated to an agent at a known cost per call.

Agent capability

AI Dashboards are now live in Docyt: a user describes a report in plain language, and Docyt AI writes it, test runs it across the portfolio and checks its totals against the P&L. A dashboard whose totals do not match the P&L is not saved. Saved dashboards run on the current books each time they open, with filters by property, account, vendor, class or department and CSV export.

Finance teams running many properties, starting with hotel groups, can get cross property reports such as DSCR against a covenant line without a data warehouse or an analyst queue. The tie out rule limits the risk of an AI built report disagreeing with the ledger.

Integrations

WhatsApp is now a native channel for Level AI Virtual Agent, connected to a WhatsApp Business Account through Meta's Cloud API with a Phone Number ID, an access token and a webhook URL, or deployed on top of an existing WhatsApp setup in a platform such as Genesys. The same agent that handles voice and web chat resolves requests there, using quick reply buttons, list menus and product or order cards, and keeps the conversation context when a customer steps away.

Teams can add WhatsApp self service without a separate bot vendor or a custom webhook build, and WhatsApp conversations flow into the same QA program as voice and chat.

Observability / auditability

The AI Gateway from Wonderful is generally available as the control layer of its AI OS, sitting between a company's AI tools and its model providers so every model call from coding, voice and back office agents passes through one point. It reports spend by department, team, user, source, task and model with budget projections and anomaly flags, enforces per team model access and spending limits before a request reaches the model, and redacts customer data before it leaves. A router picks the most cost effective approved model for each request and moves traffic to a fallback when a provider goes down.

Enterprises running agents from several vendors get one record that traces each request to a model, a cost and the reason that model was chosen, plus policy that applies whichever tool sent the call.

Human approval / guardrails

Meeting chat in Brev v5.18.14 is now redacted for contact details, pay figures and other sensitive remarks, both on the Transcript tab and before the AI writes the meeting summary, chapters and effectiveness score. A chat message Brev cannot redact is left out rather than shown, extending the privacy already applied to transcripts.

Teams using a meeting agent get a fail closed control on what the AI reads from chat, which lowers the risk of personal or pay data surfacing in AI written recaps.

Integrations

Token Security customers can now install a Token plugin from the Grok Build marketplace, so developers and the Grok Build agent itself can query Token data from the editor, for example what a GitHub token can reach and who owns it. Token also reads the Grok Build Admin API to see which users are enabled and which MCP servers and tools each team has set up, and combines that with EDR data to find what each person actually runs.

Security teams can route credential findings for coding agents to the developer who owns them, inside the editor. The plugin is read only and returns metadata about secrets, never their values.

MCP / tool calling / API

Extensions are now generally available for every incident.io Investigations customer, bringing the skills engineers have written and the tools their team has built, including internal MCP servers, into Nexus so investigations and the @incident agent can use them. Every tool is classified as read, write or destructive, and teams decide where each one can run.

Teams can point the incident agent at homegrown systems without waiting for a native integration, while per tool classification keeps control over what an investigation may change.

MCP / tool calling / API

An AI agent can now start a Stripe Atlas company formation application through Stripe's MCP server, or from the terminal with the Atlas commands in the Stripe tools plugin, for a Delaware C corporation or LLC. The agent confirms each choice with the user before recording it, and Atlas saves progress so the user can move between the agent, the terminal and the Stripe Dashboard.

Founders can start incorporation from the coding agent they already use, with confirmation before each write and identity, payment and signing steps still completed by a person.

Agent capability

Alyx, the agent built into Arize AX, can now create, edit, explain and run Decision Judge evaluators from chat. It reuses a matching evaluator or proposes a new one, sets up a project task, checks the column mappings and runs it over historical LLM spans, and nothing saves until the user accepts it.

Teams can go from a plain language description of a check to scored production traces without building evaluator templates by hand, with every saved change approved by a person and kept as a new version.

Integrations

Boomi Connect, Boomi's MCP connector service, added eight MCP connectors: Azure SQL Server, Boomi Platform, Gainsight CS, GitHub, Gong, Planning Analytics (TM1), NetSuite and Zoom Meetings.

Agents in MCP clients can reach ERP, code, customer success, sales call and meeting systems through Boomi's managed connectors instead of separately built connections.

MCP / tool calling / API

Exa 3.0 brings a larger embedding model, makes Dynamic Highlights generally available at no extra charge, and adds 1.27 billion ecommerce listings, more than 211 million local businesses, real time news, sports, weather and market data, and flights, events, transit and podcasts to its index. The release also adds Exa Lite, a lower cost search type at $4 per 1,000 requests.

Agents built on Exa can answer shopping, travel and live data questions from one search API, and highlights use 95% fewer tokens than full page text by Exa's measure. The claim that it ranks first on every benchmark tested is Exa's own.

Workflow orchestration

The new Disclosure Management Suite in Intelligize+ AI adds Custom Workflow Builder, which turns a research objective written in plain language into an editable multistep workflow that applies the same analysis across dozens or hundreds of SEC filings and can be saved and rerun. The suite also lets teams upload their own drafts to Protégé Vault and adds a Disclosure Analyzer that flags 10-K and 10-Q language to add, remove or revise, with sources and suggested wording.

SEC reporting and legal teams can move AI work from research into reviewing their own filing drafts, with each finding tied to its sources and the professional making the final call.

Agent capability

Socrates Builder, Torq's building agent, can now draft Guidance and Rules for Auto Triage from a plain English request or an attached playbook or procedure. Each drafted entry is reviewed by a person before it is enabled.

SOC teams can turn existing runbooks into the policies that steer the triage agent much faster, while a person still decides what takes effect.

Agent capability

A rebuilt Moveworks Assistant experience in Slack began rolling out on October 19 to early customers, with sandbox access for everyone else and general availability set for October 29. Answers stream in, a collapsible plan shows the steps the Assistant is taking, and admins can add Work Objects ticket cards that employees act on inside Slack and per channel limits on which plugins run.

Slack first organizations get visible agent steps and in chat ticket actions, and admins gain per channel control over what the Assistant may run. Microsoft Teams is unchanged.

Human approval / guardrails

Agentic Batch Changes now honors administrator defined context filters that make matched repositories and file paths invisible to its agents. The coordination agent cannot search or read excluded content, workspaces in excluded repositories are not run, and publishing from an excluded repository is refused, with the same rules applying to Deep Search.

Security and platform teams can keep secrets, production config or restricted repositories out of reach of an autonomous code change agent that works across many repositories.

Human approval / guardrails

Version 5.0.0 of the NeuBird MCP server turns off by default the nine tools that create, change or delete a tenant's configuration, covering connections, projects and skills. They are not listed, and a direct call is refused unless NEUBIRD_ENABLE_CONFIG_TOOLS is set to true. Every tool now carries MCP annotations that mark it read only or destructive, so clients can run read only tools without asking and prompt before destructive ones.

A coding agent connected to NeuBird over MCP can no longer reconfigure production connections or delete projects without an explicit opt in. Teams that script project setup over MCP must set the flag before upgrading, since this is a breaking change.

Agent capability

Pipecat 1.13.0 lets client apps send files and images to a running bot through a new RTVI send-file message, and the bot adds them to the LLM context. A new FileResolver downloads file URLs only from publicly routable or allowed networks, refuses redirects and caps files at 50 MB. The same release adds, in alpha, a frontend model that keeps talking to the caller while a backend model works through longer tool chains.

A voice agent can now take a document or photo from the caller mid conversation and reason over it, which opens support and intake flows that need more than speech. The download safeguards matter for teams that must control what a bot fetches.

Memory / state

Skill Workshop 2.0 in OpenClaw v2026.10.5 saves the procedures the assistant learns directly as skills, with change history, up to ten saved versions, version comparisons, archive and restore, and Undo. The upgrade retires the earlier proposal modes, in which learned skills waited for review: Auto stays on where it was set, legacy proposal modes move to Off, and /learn still works on request.

The agent now rewrites its own reusable procedures without an approval step first, so oversight moves from approving proposals to reviewing versions and undoing changes. Background learning also sends conversation and tool content to the configured model provider and can add usage charges, which teams should weigh before leaving it on.

Workflow orchestration

Moderne CLI v4.11.0 adds a factory command for Claude Code, Codex, Copilot, Cursor and other coding agents, which works across a Moderne organization by putting agent teams on changesets. A factory can start without a set objective, is hinted to spend teams sparingly, and approves its own tool use without asking.

This moves Moderne from one prompted agent toward an unattended multi agent loop across a code estate, so buyers should decide where self approved tool use is acceptable before turning it on.

Agent capability

Sandboxes give an agent its own isolated Linux machine with its own CPU, memory and disk, started from a managed image such as Debian, Node.js or Python or built from a Dockerfile. Commands and files run through the CLI or the TypeScript and Python SDKs, disks can outlive a machine, idle machines can pause or hibernate, and outbound firewalls control what a machine can reach.

Teams that already route agent tool calls through Metorial can give those agents a place to run code and keep files without adding a separate sandbox vendor, with per machine isolation and egress controls a security review will ask about.

Agent capability

Five new Intelligent Agents from Enboarder join managers, buddies, HR and IT as stakeholders in employee journeys. Skills Coaching, Knowledge Check and Role Play prepare new hires through AI chat, including rehearsing conversations with the AI playing a customer or patient, while Check-In and Activity Reflection run adaptive conversations and summarize them for the manager and HR. A Manager Mode in the Lifecycle Assistant and an enhanced manager dashboard put that output in front of the person who has to act on it.

HR teams get conversational training, assessment and pulse checks inside the onboarding journeys they already run, in place of static forms and quizzes, with decisions about people left to the manager.

Human approval / guardrails

Spoki v19.28.0 adds a Behavior tab to each AI agent, where admins describe the situations in which the agent must hand a conversation to a person, such as refund requests or complaints. Previously the agent transferred only when the customer asked for a person or when it could not help. After a handoff, the agent pauses its replies for 24 hours so the human team keeps control.

Teams running WhatsApp agents can set escalation policy agent by agent, which limits where the AI acts alone on sensitive topics and gives staff an uninterrupted takeover.

Agent capability

Narada Code is a computer use agent harness that investigates the applications involved, drafts an enterprise workflow, tests it in a sandbox and revises it after failed runs, extending coding harnesses such as OpenAI Codex and Claude Code. OperatorMini, launched alongside it, is a lighter execution agent that uses 6 to 8 times fewer credits than Operator. With workflows built by Narada Code, Narada reports Operator at 86.1% on WebArena and OperatorMini at 83.5%.

Building and repairing browser and desktop automations usually falls to engineers, and this hands the test and repair loop to an agent with a cheaper executor for repetitive runs. The WebArena figures are vendor reported.

Human approval / guardrails

Version 1.11.0 lets a NudgeBee automation be set to run without asking for confirmation. The same release blocks irreversible cloud operations during remediation, runs the agent workspace on a read only file system, and marks automation steps the cluster agent refuses as failed.

Teams can let trusted runbooks act on their own while destructive cloud actions stay off limits, so on call engineers approve fewer routine fixes.

Agent capability

TREX, Greptile's end to end testing agent, is now generally available. On each pull request it starts the app in a sandbox, clicks through the flows the change affects, and posts what it found with a captioned video of each flow, and the release adds environment setup with network access controls, a simulation mode for testing that setup, and custom flows.

Pull requests can get runtime testing alongside code review at a published 3 credits a run, and setting network access for the sandbox lets security teams limit what it can reach.

MCP / tool calling / API

A new workflow builder API can create and edit any workflow the OmniDimension dashboard can build. Developers read the catalog of node types, apply node additions, changes, moves, connections and removals in one batch that applies fully or not at all, and validate a workflow before publishing it, and the Python SDK and the OmniDimension MCP server support it too.

Voice agent workflows can now be generated and changed from code or from an AI assistant over MCP, and the all or nothing batches plus validation lower the risk of a half applied edit reaching live calls.

Human approval / guardrails

Release 1.55.0 of goose binds subagents to their parent's hooks, goose mode and tool approvals, so work the agent delegates runs under the same checks as the parent session. The agent also stops inferring or steering while a tool approval is pending. The same release removes background subagents and the orchestrator and deletes the legacy agent loop.

Teams that rely on approval modes or hook based policy scripts now get the same controls on delegated work. Anyone using background subagents must rework those flows before upgrading.

Human approval / guardrails

The Agent Control Plane is now generally available on CrewAI AMP, giving platform administrators 13 built in policies to configure once and switch on across deployments, covering PII detection and masking, allowed models and tool blocklists, execution windows, safety and compliance rules, and cost limit alerts. Teams can add custom policies that evaluate at runtime without code edits or redeploys, and the policies apply whether an agent was built in Crew Studio or in code.

Security and platform teams get one place to restrict which models and tools any AMP automation may use and when it may run. Cost limits notify rather than stop a run, and CrewAI's platform docs still describe the control plane as beta, so confirm the full policy catalog in a demo.

Agent capability

The Playcode SDK now gives apps built on Playcode an AI chat agent with each user's conversations saved in the app's database, picking up from the last finished step after a reload or restart. Actions a user must confirm, such as a booking or an email in their name, wait for Approve or Deny in the chat, and actions only the owner allows, such as a refund, wait on the app's admin page. Owners can also set a daily spending limit per user and name a fallback model for when the first model's maker is down.

Approval gates and per user spend limits arrive as defaults rather than custom code, which lowers the risk of putting an agent that takes actions and spends model credits in front of customers.

Integrations

The mid October StackAI release adds a broad set of integrations, among them Akeyless and Keeper secrets management with Keeper SCIM provisioning, Chatwoot, Jamf Pro and Egnyte file triggers, plus an action to approve GitHub pull requests. Salesforce, Coupa and QuickBooks MCP connections can now use StackAI managed OAuth without registering an app, and custom MCP connections accept a custom header name for API key authentication. Workflow, Agent Builder and Interface also now share one autosaving draft engine with version history.

Managed OAuth removes the app registration step that often stalls MCP rollouts with core finance and CRM systems, and the shared draft engine gives teams a clearer record of which revision went live.

Agent capability

NexusAI, the new agentic experience inside ThreatModeler Nexus, builds a threat model from a diagram, design document, screenshot, infrastructure as code file, repository or plain description, asks about missing details and matches components to the customer's libraries. It recommends which threats and requirements apply with evidence strength and confidence signals, and the team decides which recommendations to apply. Scheduled daily or weekly checks of GitHub, GitLab, Google Workspace and Jira or Confluence through MCP detect architectural drift and stage a model update for review.

Security teams can hand model creation and first pass analysis to the agent while keeping the decision step, with each applied change recorded. The launch post does not state availability terms, so confirm access in a demo.

Agent capability

AI Voice Agents from Aircall can now act on Shopify orders instead of only looking them up, checking returns and refunds, holding a shipment, taking an address change request, answering product questions and sending a payment link. Cancelling orders and issuing refunds are sensitive actions that are off by default, turned on through the Aircall account team and reachable only after the caller reads back a one time code texted to the phone number on the customer record. Every change the agent makes is noted on the Shopify order with an "Aircall AI:" prefix.

Ecommerce support teams can resolve post purchase calls on the phone line while money moving actions stay gated behind caller verification. Codes go by text message only for now, so callers found without a phone number cannot reach the sensitive actions.

Integrations

An official Allo plugin for Grok Bot lets a bot run the Allo phone system after a single Allo sign in, with no API key. The bot can set up lines (who rings, opening hours, what callers hear and the AI receptionist), report on calls and who to call back first, find what a customer said, text customers and line up the Power Dialer. It checks with the user before anything reaches a customer, and line changes are published in Allo.

Teams that work in Grok Bot can manage calls and follow ups from chat, with a confirmation step before any message reaches a customer.

Memory / state

Release 1.6.5 of Cognee adds an opt in fact conflict review to improve(), replacing its earlier contradiction detection, and a BROAD search type for questions about a whole dataset. The release also adds a Google Calendar connector, turns raw Cypher search off by default and deprecates memify in favor of improve().

Agents that read stale or contradictory facts give wrong answers, so a review step for conflicting facts gives builders a way to catch memory drift. Teams using memify or raw Cypher search need to plan for the deprecation and the new default.

Pricing / packaging

AI Builder, formerly Hostinger Horizons, now draws on one shared AI credit balance with Hostinger Agent, Hostinger Reach and GPU instances, fed by subscriptions, hosting plans, top ups, gifts and free trials, with trial credits used first. AI Builder credits moved to a 50 to 1 scale, so one former credit equals 50 AI credits at the same value, and hosting plans now include 250 to 1,000 AI credits.

Buyers budget AI use from one pool across several Hostinger products, and old AI Builder credit figures need converting at 50 to 1 when comparing plans.

Human approval / guardrails

Akka SDK 3.7.0 adds a new guardrail API: a guardrail runs before each model call, on the agent's final reply, or before a function or MCP tool call, and allows or denies it. The same release adds Evaluators, which run in the background after an agent's interactions and record a verdict in a ledger that teams can query, and it deprecates the older TextGuardrail and SimilarityGuard APIs.

Teams can block an agent's tool and MCP actions in runtime configuration rather than relying on each developer to add checks, and the ledger gives auditors a record of what each agent did and how it was scored.

Workflow orchestration

Capy 0.4.7 lets a Captain, the agent that coordinates a crew of other threads, run on one of the user's paired devices, and the Captain can start crew threads on any paired device or in the cloud. A Captain can also go back to being a regular thread.

Multi agent work led by a Captain can now run on a company's own machines rather than only on Capy's cloud machines, which matters for code and data that should stay on local hardware.

Agent capability

Gitar, Sonar's code review agent, added Agentic Security, which reviews changes to agent skills together with the scripts they call to catch risks such as a removed approval step or leaked credentials, drawing on Sonar's collaboration with OpenAI on Daybreak. The same release adds Cross-PR Analysis, which flags open pull requests that could break each other within or across connected repositories, on by default for Enterprise.

Teams whose coding agents edit their own skills get a review step aimed at the changes that loosen controls, and conflicting pull requests surface before both are merged.

Agent capability

In careCycle 2.0 every call opens with a member fact sheet listing policies with their writing agent, appointments with reference codes, open requests and whether anyone acted on them, and summaries of the last eight calls. The assistant can now move or cancel an appointment during the call using the same steps as the operator reschedule and cancel paths, and on the problem calls careCycle tested, repeat notes fell from 14 to 1 and false promises from 33 to 6.

Agencies can let the assistant finish appointment changes on the call instead of routing them to staff, and starting from full member history should mean fewer duplicate requests and unkept callback promises. The figures come from careCycle's own test calls.

Deployment / data residency

The agentic conversational platform from Gnani.ai is now live as a fully colocated stack in the US, Singapore, Japan and the UAE, with production voice conversations already running. In each region, orchestration, inference, conversation data, tool calls and observability run in the country across voice, chat, text and email, offered as a managed in region deployment or dedicated single tenant infrastructure.

Regulated buyers in these four countries can run voice and chat agents without conversation data leaving the country during processing, not only at rest. The Philippines is named as the next region but is not live yet.

Deployment / data residency

Agno v3.1.4 adds a sandbox parameter so ClaudeAgent and CodexAgent run their whole harness process inside an isolated computer for each session, with DockerSandbox as a provider the customer operates and AgentCoreSandbox on Amazon Bedrock AgentCore Runtime. AgentOS accepts foreground, streaming and background runs for sandboxed agents and keeps a session's computer warm for 300 seconds by default.

Teams that put Claude Code or Codex behind their own product can confine the coding agent to a Docker host they run or to AgentCore rather than the AgentOS server, which limits what a misbehaving agent can reach.

Workflow orchestration

Cline for VS Code v4.1.25 starts rolling out Cline Cloud sessions, which run a task in a hosted sandbox on a GitHub repository and branch the user picks. Cloud tasks show in Recent and History with their status, notify the user when they finish, and keep running after the editor closes.

Developers can hand long tasks to isolated cloud sandboxes from the editor without keeping a laptop open, which keeps agent execution off developer machines.

Deployment / data residency

Junie Local, the fully on device version of the Junie coding agent, moved from experimental to stable on Windows PCs with NVIDIA RTX GPUs of at least 24 GB and added native support for NVIDIA RTX Spark PCs on Windows on Arm. The rebuilt installer checks the GPU and driver, sizes the context window to available GPU memory, and offers Qwen3.6-27B or a JetBrains blend of it.

Teams that cannot send code to a cloud model can run a coding agent entirely on standard Windows workstations, with no code leaving the machine.

Security / enterprise

Project API keys in Langfuse can now be created with a permission level: Full access, Viewer, Ingestion (create traces, scores and media only) or Scores ingestion. Existing keys keep their current access.

An ingestion only key embedded in an agent or web app can no longer read back traces, prompts or datasets, so a leaked key exposes far less of the project.

Security / enterprise

Leo conversations with the open weight GLM model are now end to end encrypted. The browser encrypts each prompt on the device so that only a trusted execution environment that has passed hardware attestation can read it, and Brave says the setup keeps prompts and responses from infrastructure operators, including Brave itself. It is available in Brave's Release, Beta and Nightly channels.

Anyone sharing sensitive pages or documents with a browser assistant gets privacy they can verify rather than a policy promise, and the release extends the confidential computing setup Brave announced in November 2025.

Deployment / data residency

New Amazon EC2 images for Crogl release 2.12.1 include an air gapped image that boots with its egress switch already on. With the switch on, the host and every agent container can reach only private ranges, the instance metadata service, Amazon Time Sync and an allow list, and the switch also stops the standard Ubuntu update and background services that would otherwise contact public endpoints.

Regulated and classified SOCs can launch an instance that is closed to the internet from first boot instead of hardening a connected image themselves, which shortens the accreditation work for a sovereign deployment on AWS.

Browser/computer use

Version 3.1.0 of the Bright Data MCP server turns every tool on by default, so the Agent Browser tools no longer need to be enabled separately. Results from the browser, search, scrape and dataset tools are now marked as untrusted web content, so the model treats text on a visited page as data rather than instructions.

Any agent connected to the server can drive Bright Data's remote browser with no extra setup, and the untrusted content marking is a first defense against prompt injection from the pages it opens.

Human approval / guardrails

OCP Release 3.7.0 lets individual agent tasks override the default retry counts for human agent requests and for no input, so a voice agent can hand a caller to a person sooner when they ask for one. The release is rolling out, with some services still completing deployment.

Contact center teams can set escalation thresholds task by task instead of platform wide, which gives finer control over when the agent stops trying and a person takes over.

MCP / tool calling / API

Every Signals deployment now serves an MCP server at /mcp on its Signals API URL, with nothing to install and authentication through the existing Snowplow API key. Four read only tools let an agent in the customer's own product read calculated attributes, up to 25 attribute groups per call, and recent raw events for the current user.

Teams building customer facing agents can give them live behavioral context for each user without writing retrieval code for every data set.

Memory / state

Action Index stores web workflows that have already succeeded, keyed by website as parameterized actions, and replays them as deterministic code, calling the model only at steps where the page has changed. It is available to every Anchor account, including the free tier, inside web tasks, through the REST API and as tools on the Anchor MCP server, and a per site MCP server exposes each active action for one site as a typed tool. On Anchor's own Online-Mind2Web run of 300 tasks, gpt-5.6-luna rose from 62.7% to 93.3% success with Action Index, at $0.016 per task.

Agents that repeat the same portal and form flows stop paying for full inference on every run, so cost per completed task falls and smaller or open weight models become workable for known flows. The figures come from Anchor's own benchmark runs, not customer workloads.

Integrations

GitLab.com and self hosted GitLab instances can now connect to Corelayer over OAuth, with the customer choosing which projects the agent can use for code context. Indexing the code and opening merge requests are optional settings, and both are off by default. The same release lets teams file an investigation as a linked Jira or Linear ticket carrying its summary, root cause and next steps.

Teams on GitLab, including self hosted instances common in regulated industries, can give the investigation agent code context while write access stays opt in. Root cause findings also land in the existing ticket queue without copy and paste.

Observability / auditability

Event Auditing, new in Itential Platform 6.6, is an immutable audit trail that records who performed an action, what they changed and when, and actions taken by FlowAI agents are written to the same trail as actions taken by users. It ships in 6.6.1, released October 15 to replace the 6.6.0 build, which is no longer available. The release also sends web server logs to Grafana over OpenTelemetry, and its changelog adds audit logging for authorization, service and Studio resources.

Compliance and security teams get one record of changes made by people and by FlowAI agents, which is the evidence auditors ask for before agents touch production network infrastructure.

Integrations

A Bring Your Own Agent integration lets teams build monday.com agents whose logic runs in a Make scenario, and people can chat with the agent, @mention it on an item or assign it work like a teammate. It adds four beta Make modules: Watch Agent Events, Send Agent Response, Run monday.com AI and Run Platform Agent. The agent reads and writes board data through the hosted monday MCP server, using the same connection as the agent.

Teams that already automate in Make can put their own agent logic behind a monday.com agent identity without building a custom app. The modules are in beta.

Human approval / guardrails

Zenity now supports tool call validation in Anthropic's Inference Hooks for Claude Enterprise, so each tool call Claude proposes is held for an allow or deny verdict from Zenity before it runs. Zenity Boundaries can block write, delete and execute operations, calls to MCP servers outside a reviewed list, and actions traced to prompt injection, with one configuration covering claude.ai, Cowork, Claude Code, Claude in Chrome, Claude for Microsoft 365 and Claude Tag. Nothing is installed on endpoints, and customers ask their account team to turn on enforcement.

Security teams can approve write capable connectors, MCP servers and wider Claude Code rollouts with policy on each action, instead of finding violations after the fact. One denial stops every tool call in the same response, so policies need tuning to avoid blocking permitted calls.

Agent capability

Figma Make now scaffolds a Payload backend natively, giving an app built in Make a CMS, database, media storage, APIs, authentication and an admin panel, so a prototype can become a production app. Figma hosts the backend, so content and data live alongside the designs that shape them. It is rolling out across all plans and was announced with other Make updates at Config India.

Teams can take a Make prototype to a working, data backed app without standing up a separate CMS or database. Hosting with Figma also means app data sits with Figma, so residency and handoff needs deserve a check first.

Agent capability

Datarails Spend, now available on the FinanceOS platform, brings purchase requests checked against the live budget, purchase orders, contracts, invoices, corporate cards, bill payments and ERP reconciliation into one system. AI handles invoice extraction, coding, transaction classification, matching invoices to purchase orders and contracts, and flagging discrepancies and unexpected charges for review. Approved bills are scheduled for payment before any money goes out.

Finance teams get AP coding and invoice matching done by AI against the same budget lines they plan on, while approval of requests and payments stays with people. Actual spend flows back to the budget line it was approved against.

Integrations

Google Sheets and Microsoft Excel apps arrived in Nextiva's Integrations marketplace, and once connected, the XBert AI Employee can look up and add rows in a spreadsheet during a call or chat. Both apps also cover totals, charts, new trackers and PDF export, and Excel can update ranges and formulas.

A small business can put a price list, stock sheet or callback list it already keeps to work in live conversations, without a CRM project.

Integrations

Slack is now an Agent connector in Pigment, alongside Notion, giving Pigment Agents access to Slack content from inside Pigment. According to Pigment's connector setup guide, admins enable it for the workspace with credentials from a Slack app they create, and set each read or write tool to Always allow, Needs approval or Blocked.

Planning context that lives in Slack threads can reach the agents without copy and paste, while the per tool settings keep Slack write actions behind a person's confirmation.

Human approval / guardrails

Qevlar AI now prepares a response for every new alert that joins an alert pattern an administrator or manager has activated. In Supervised mode nothing runs until an analyst approves, in Automated mode actions go straight to the customer's SOAR through an automation endpoint, and responses are always held when an alert involves a VIP user or has no safe target. Qevlar does not execute actions itself, and every decision, by Qevlar or an analyst, is written to the alert's audit log.

The AI SOC moves from recommending next steps to driving response on repeat alert patterns, under a mode administrators set per pattern. Buyers need a SOAR or workflow tool in place to act on it.

Agent capability

Rippling relaunched Automated Compliance with multi framework support (SOC 2, ISO 27001 and HIPAA today), AI help setting up templated or custom frameworks or importing an existing program, and evidence collection from connected third party systems as well as Rippling's own HR and IT tools. It adds AI suggested remediations that admins can carry out, a built in audit portal, publishable trust centers and AI help with security questionnaires.

GRC teams can run several frameworks and close control gaps from the system that already holds HR, identity and device data, and Rippling positions it for mature programs whether or not the customer uses its other products. ISO 27701, ISO 42001, GDPR and PCI DSS templates are slated for December.

Agent capability

Glean release 563 adds Skill reflection, which picks out the feedback users give in chat while running a Skill and proposes it as an edit to the Skill's instructions. A user can also type /reflect at any time, reviews the feedback behind each proposal and approves it before it applies, and editors of shared Skills can improve them the same way.

Reusable agent Skills can improve from everyday corrections without someone rewriting them by hand, while a person still signs off on every change so one off requests do not become permanent rules.

Memory / state

Haystack 3.4.0 adds an experimental ToolResultOffloadCompactor for the Agent's compaction hook: older tool results are written to a ToolResultStore and replaced by a reference, and the agent can read the full output back on demand through a read tool scoped to the same store. The release also adds Python 3.15 support and raises the minimum to Python 3.11.

Agents that call tools with large outputs can stay inside the context window without dropping or summarizing those results, and teams still on Python 3.10 must upgrade Python before moving to 3.4.

Pricing / packaging

The new Nevermined Subscription lets buyers pay once by card for an allowance of credits and spend them across the AI Services Catalog, instead of paying per call in crypto. An agent or MCP host connected with an API key spends credits automatically, bounded by the prepaid balance.

Prepaid credits cap agent spend at the balance and remove per call crypto signing, which opens the catalog to teams whose finance rules rule out holding wallets.

Human approval / guardrails

Agentic Commerce Defense in Operant's Agent Protector evaluates each agent payment attempt over the x402, MPP and AP2 protocols before it is forwarded. The decision weighs the agent's task, the paid tool call, the MCP server, the payment destination, the amount and the organization's commercial policy.

Teams letting agents pay for APIs, MCP tools or goods get a per transaction check that can stop an injected or off task payment even when the credentials are valid and the amount is under budget.

Pricing / packaging

DBOS Transact for Python 3.3.0 moves the DBOS Conductor client out of the open source library into a separate Enterprise package. Applications that use Conductor or DBOS Cloud must install the Enterprise dependency when upgrading, while users of only the open source library need no changes.

Teams running durable agent workflows with Conductor or DBOS Cloud have a required step when they upgrade, and the change separates the open source library from DBOS's enterprise components.

Agent capability

Code Review, the Claude Code agent that reviews GitHub pull requests, checks its own findings and comments inline, is out of research preview and generally available to Claude Team and Enterprise organizations. Admins can set Quick, Standard or Deep review levels for the organization or per repository, and Anthropic says most reviews now finish in 6 to 10 minutes instead of about 20.

Teams can run automated pull request review on a supported product and trade depth against cost repository by repository. Eligible Enterprise organizations can claim at least $10,000 in Code Review credits by January 15, 2027.

MCP / tool calling / API

Tool Recommendation, new in Arcade, matches each task to the few tools an agent needs at that moment instead of loading Arcade's whole catalog of more than 8,000 tools into context. It is built into the Arcade actions runtime, so governance and policy enforcement still apply to every recommended call.

Teams can give agents the whole tool catalog without maintaining a tool list per agent. Arcade's own figures claim up to 91% fewer input tokens per session and up to 37% lower token costs.

Observability / auditability

Analytics in Consio now has an AI agents page reporting calls handled by AI, automation rate, resolution rate, average handle time and revenue from AI handled calls. It breaks calls down by intent with the AI resolution rate for each and shows which calls passed to a human.

Teams running Consio voice agents can see where the agent resolves calls and where it hands off, which shows what to tune before widening its scope.

Workflow orchestration

Legal Work Orchestration organizes a matter in Harvey into phases and tasks, each carrying its scope, inputs and reviewer, so agents can execute defined tasks and lawyers review the output at designated gates. Harvey notifies the reviewer when an agent finishes, moves to the next step on approval, and lets lawyers close or start tasks by replying to email, while reusable matter templates are in early access.

Harvey moves from assisting prompt by prompt to agents running defined steps of a matter under explicit approval gates, the structure firms need before delegating more work to agents.

Pricing / packaging

HubSpot customers can now move spend from unassigned seats to HubSpot Credits on their own at any time, instead of waiting for a contract renewal. Credits are the units that Breeze agents such as Customer Agent and Prospecting Agent consume, and the same post covers account level credit limits and pay as you go overages.

Teams expanding agent work partway through a contract can shift budget from idle seats to agent usage without renegotiating, while keeping a ceiling on what agents can spend each month.

Agent capability

Agent accounts in Copilot Studio let agents work in Microsoft 365 with their own credentials, so people can assign them work in Teams and Outlook, mention them in Word, Excel and PowerPoint comments, and hand off tasks the agent continues between interactions. Tools run under the agent's own approved credentials and scoped permissions rather than a user's or maker's, and agents keep business process memory across the people they work with.

Agent actions become attributable to a governed identity that admins can scope and manage, which matters for processes that outlive any one employee. Microsoft says more access policy, cost and lifecycle controls are planned.

Agent capability

Phonic's Python SDK v0.32.39 and TypeScript SDK 0.32.33 add web_search to the built in tools a voice agent can be given, alongside keypad input and natural conversation ending. The builder can choose whether the agent speaks before it runs a search mid call.

A voice agent that can look things up during a call can answer questions outside its prompt without the builder hosting a search webhook. The docs did not yet describe the tool, so behavior and cost are worth confirming with Phonic.

Security / enterprise

AI Supply Chain Security 26.10.0 lets a tenant add multiple rules to its Supply Chain policy. Rules can be built in a Rule Builder from attributes, values and operators, or picked from preconfigured rule templates, and they decide how scanned models are judged against the organization's policy.

Security teams that gate which models reach production can encode their own admission criteria in the product, which helps cut noise in pipelines that block noncompliant models.

MCP / tool calling / API

Zep's v4 API is generally available, with v4.0.0 SDKs for Python, TypeScript and Go, and v4 is now the default documentation. Users, threads and graphs are addressed by Zep generated UUIDs, every error response carries a stable code, and the SDKs send an idempotency key on retried creates so a retry does not make a second resource.

Agents that call Zep while they work can branch on error codes and retry safely. Moving from v3 user and thread IDs to Zep UUIDs is a breaking change teams on v3 need to plan for.

Workflow orchestration

Restate Server v1.8.0 turns on Virtual Queues, the scheduler behind its flow control features, for every cluster, so scopes, limit keys and scoped Virtual Object targets work without opting in. The same release rebuilds the restate CLI for scripts, CI and AI agents, with JSON output and a preview step before state changing commands.

Teams running agents for many tenants can cap concurrent runs per tenant, which bounds model spend and protects downstream APIs. The upgrade needs care, since nodes must first run v1.7.3 or newer and each partition is briefly unavailable while it migrates.

Integrations

The AI Receptionist and Virtual Receptionists from Smith.ai now connect directly to a firm's Lawmatics account for scheduling and CRM. On the call, the receptionist qualifies the caller, checks the connected Lawmatics calendar and books the consultation, and afterwards new leads and existing clients are logged in Lawmatics with the call summary attached.

Law firms running intake on Lawmatics get consultations booked during the call, before a lead can ring another firm, and intake data lands in the CRM without manual entry.

Integrations

Meta's Muse Spark is now a tracked platform in Writesonic's AI visibility monitoring. Users can select Muse Spark in the Platform filter to review its answers to their tracked prompts, with visibility, position, sentiment, brand mentions, citation status and the pages it cites.

Personal agents like Muse research and shortlist products for a user without a site visit showing up in web analytics, and marketing teams can now see how that agent positions them next to the engines Writesonic already tracks.

Agent capability

Version 1 of Kerno's runtime QA agent for backend code, its largest release so far, tests whole user flows such as signing up, creating a project and inviting a teammate against real databases, queues and services, and reruns every flow a code change touches. It shows a test plan for approval before writing tests, checks each entry point against the OWASP API and Web Top 10 categories that apply, and tests MCP server tools at runtime. A rebuilt portal adds an organization view of flows covered, code changes tested and issues caught, with a heatmap of how tests changed day by day.

Teams whose coding agents ship backend changes get runtime evidence inside Claude Code, Cursor or Codex before a merge, and engineering leaders get one record of what the testing agent checked across the organization.

Integrations

Microsoft Teams calls can now be captured in Salesloft Conversation Intelligence by importing Teams cloud recordings through the Microsoft Graph API, instead of sending a bot into the meeting. Admins set the Teams recording method to Cloud Recorder, turn on auto record and choose which recordings to import. The method has no live transcription or real time features, and calls take longer to appear because recordings are checked every 15 minutes.

Organizations that block meeting bots can now get Teams calls captured, summarized and searchable in Salesloft. Setup needs a Microsoft admin to grant organization level consent to the Graph API.

Observability / auditability

A new Variables tab in Synthflow's call logs shows the value each variable held on a specific call, covering system variables, variables set before the call, collected variables and inbound SIP headers. It also lists every action's inputs and outputs on that call, including transfer SIP headers.

Teams debugging a voice agent can see exactly what data it sent to and received from each action on a call, instead of piecing it together from transcripts and webhook payloads.

MCP / tool calling / API

Version 5.8 of the Blits.ai platform lets customers publish their agents over the Agent2Agent (A2A) protocol, so other platforms and agents can discover and talk to them. Each publication gets a public agent card for discovery with JSON RPC and HTTP bindings, protected by an API key, and nothing is discoverable until it is published. The same release adds a per conversation decision log of routing decisions, tool calls and guardrail verdicts that can be exported to a SIEM.

Agents built on Blits.ai can be called by agents on other vendors' platforms without custom integration work, and unpublishing takes an agent offline immediately, which gives security teams a clear switch for external exposure.

Security / enterprise

Release v2026.10.14 adds data access controls for Pluvo admins: a member can be granted a whole data source or specific tables, then narrowed by hiding tables or columns or adding a row rule, with a log of recent changes. New non admin members now start with no access to synced data until an admin grants it, while existing members keep their current access. The same release renames Jobs to Agents across the product.

Finance teams can give department heads access without exposing payroll, bank details or other entities, since synced data can now be scoped per person down to the row.

Workflow orchestration

In GWork V1.0.32, the GPTBots workspace app, a user can mention task B with @ inside task A, and the agent pulls information from B or assigns work to B and receives the results. Task B shows the record too, with states that mark whether it is running or idle. The same release lets a project give the agent access to several folder paths and lets the agent read, summarize and reply to email from several work accounts in the new My Space.

People running several agent tasks at once can hand work between them inside the app instead of copying context by hand, which brings multi agent coordination to end users rather than only developers.

Human approval / guardrails

Authentication Gate, new in the Talkdesk AI Agent Platform, blocks access to sensitive Sub Agents and Skills until the customer session is authenticated. Admins define the authentication conditions with fixed or computed variables and apply them at the Sub Agent or Skill level, and unauthenticated requests are routed to the company's own authentication flow before they reach the LLM.

Banks, insurers and healthcare contact centers can let AI agents handle account actions while identity verification is enforced in configuration rather than in prompt instructions, which lowers the risk of an agent acting for an unverified caller.

Agent capability

AI Agent on Socials is now generally available to all Gorgias AI Agent customers, answering shoppers in Instagram DMs, Facebook Messenger and WhatsApp with the same knowledge, skills, guidance, actions and settings used on other channels. It takes order actions inside the thread, recommends products with add to cart through Shopping Assistant, hands conversations to the team in the same thread and reports performance by social channel.

Brands that sell through social DMs can run the agent they already configured on those channels instead of a separate social bot, with a separate tone of voice per channel if they want one.

Observability / auditability

Conversation Intelligence and AI QA, built on Kustomer's AI Attributes, launched on October 14. Conversation Intelligence classifies, scores and extracts data from every conversation and feeds the results into routing, workflows and reporting, while AI QA scores every conversation against business defined criteria with adjustable weights and auto fail logic, showing the reasoning behind each score.

Teams running AI and human agents get quality scoring across all conversations instead of a small sample, and the same scores can be used to tune AI agents and coach human reps.

Integrations

RunSybil's October release pushes Sybil's pentest findings into Jira as issues with custom field mapping, and updates the existing ticket when a later test rediscovers a finding or a retest completes. Teams can also connect Notion so Sybil's pentesting agents draw application context from the team's knowledge base during a test.

Teams that run remediation in Jira get agent findings in their existing queue without working in the RunSybil interface, and the agents can read the documentation a team already keeps instead of being briefed for each test.

MCP / tool calling / API

Retool 4.74 adds tools to the Retool MCP server for disabling and enabling users and for adding and removing permission group members, alongside a tool to list user spaces and space filtering on the audit log tool.

AI clients connected over MCP can now offboard users and change permission group membership in Retool, so admins should review which MCP grants their agents hold.

Observability / auditability

The Analytics tab in SleekFlow AI is now one dashboard for every AI agent in a workspace, replacing a view that showed one agent at a time. It compares AI with staff and adds a date range and a comparison period.

Buyers can see how much of the conversation load AI agents carry across every agent at once, which is the core figure for judging automation coverage and staffing.

Workflow orchestration

Colleagues who can work with a shared Spinnable AI worker can now ask it in chat to create scheduled tasks and webhooks, then list, change, pause, resume or cancel the tasks they wrote, while managers can manage any task. Scheduled work runs with the writer's recorded permissions, and a webhook cannot be created with permissions above its creator's.

Teams can share one worker for recurring and triggered work without routing every setup through its manager, and permission scoping keeps a teammate from using the worker to reach data they could not reach themselves.