Human approval / guardrails

OpenAI Agents SDK v0.17.6 added pre-approval tool input guardrails.

Direct guardrail control on tool execution — one of the clearest SDK-level agent-governance changes in the window.

Observability / auditability

OpenAI Agents SDK v0.17.6 added SDK-only custom data for tool outputs.

Improves instrumentation and downstream metadata handling around tool execution, relevant for auditability and richer action traces.

Browser/computer use

Stack AI's E2B partnership now powers sandboxed Custom Code, Computer Use/RPA replay, and Browser Use on the StackAI platform.

Concrete browser/computer-use infrastructure change with enterprise security implications, especially for regulated-industry deployments backed by isolated E2B VMs.

Workflow orchestration

LangGraph 1.2.6 fixed nested subgraph checkpoint namespace inheritance and cancellation of running subgraphs on v3 stream abort.

Nested subgraphs and stream interruption behavior are core orchestration primitives for multi-step and multi-agent systems.

Agent capability

Haystack 2.30.2 prevents Agent from exiting prematurely when an invalid tool call is discarded; the model can continue looping and recover.

Improves real agent behavior in multi-step/tool-calling runs and improves recovery from malformed tool outputs.

Workflow orchestration

Cursor Automations gained /automate, Slack emoji triggers, five new GitHub triggers, default computer-use support for automation-run cloud agents, default PR opening, and memory-file deletion controls.

Materially changes the trigger surface, creation path, and action surface for always-on agents — relevant for teams building automated coding workflows.

Workflow orchestration

[Pre-release / alpha] CrewAI 1.14.8a / 1.14.8a1 introduced JSON-first crews and expanded FlowDefinition: script/code block actions, crew actions, each composite actions, expressions, DMN mode, no-Python flow run tools, human feedback from flow definitions, config/persistence wiring, experimental `crewai run --definition`, ZIP deployment fallback, and optional `if` expressions for `each.do`.

A large workflow-definition and deployment-surface expansion. This is an official alpha/pre-release — buyers should not rely on these capabilities until stable release confirmation.

MCP / tool calling / API

Tray.ai made MCP dynamic authentication generally available. MCP tools can now run using end-user credentials instead of shared service accounts, with allowlists, authentication management, and a usage monitor.

Materially improves MCP governance, identity propagation, and enterprise traceability for tool calls.

Observability / auditability

Voiceflow added transcript filtering by workflow, playbook, and tool usage.

Improves post-hoc analysis and operational debugging by linking conversations to the workflow and tool paths actually used.

Pricing / packaging

Voiceflow added organization-level usage breakdowns showing spend by workspaces, projects, date range, and by Runtime, Measure, Development, and Generation features.

Changes billing/usage visibility available to operators and improves packaging transparency for larger organizations.

Workflow orchestration

Pipecat v1.4.0 added realtime_service_mode on LLMContextAggregatorPair, the on_user_turn_message_added event, and RealtimeServiceMetadataFrame, changing how realtime speech-to-speech services write context and expose turn behavior.

Significant orchestration/runtime change for realtime agents — affects turn timing, latency, server-vs-local turn strategy selection, and pipeline metadata.

Workflow orchestration

Pipecat v1.4.0 added the pipecat create project-scaffolding CLI via an optional cli extra, including optional Pipecat Cloud enablement.

Lowers the barrier to packaging and standing up agent projects, including cloud-oriented workflows.

Funding / partnership

Legora and Ironclad announced a phased AI-to-AI integration that will let mutual customers bring Ironclad contract intelligence into Legora and bring Legora legal intelligence into Ironclad.

Materially expands enterprise system-of-record connectivity and cross-platform legal-agent workflow coverage for in-house legal teams.

Observability / auditability

deepagents-code 0.1.19 added dual-write of agent traces to extra LangSmith projects.

Concrete observability/routing improvement for teams managing multi-project LangSmith operations.

Integrations

Adomik became a default Dust integration via OAuth with read-only access to monetization analytics and the Adomik knowledge base.

Relevant for specialized vertical analytics workflows where monetization data should be agent-accessible.

Integrations

Lemlist MCP Server became available across all Dust workspaces for lead retrieval, contact enrichment, and sequence management.

Expands outbound-sales workflow coverage for teams using Dust as a workspace AI platform.

Deployment / data residency

Cursor added cloud environment setup in the Agents Window, reusable snapshots via .cursor/environment.json, /in-cloud cloud subagents on isolated VMs/branches, and local-to-cloud handoff.

Materially expands remote execution, reproducibility, parallel agent execution, and local/cloud handoff for teams running Cursor in CI or multi-agent workflows.

Workflow orchestration

Devin Playbooks now support a structured output schema so Devin can return defined JSON results.

Improves deterministic workflow outputs and downstream automation compatibility for teams integrating Devin Playbooks into automated pipelines.

MCP / tool calling / API

Devin added an official Axiom MCP server and improved MCP error cards with direct View logs links and detailed plugin status.

Adds another official MCP server and improves operational visibility for MCP failures — relevant for teams debugging agent tool-use in production.

Workflow orchestration

Devin added Slack / automation / Linear workflow improvements: preceding Slack thread context, `!agent` routing into full agent sessions, default Slack sync, automation-result posting to Slack channels, and Linear project filters for automation triggers.

Materially improves collaborative agent handoff, automation routing, and workflow context inside Slack and Linear for enterprise teams.

Security / enterprise

Devin Review now includes a Security Findings section, and the security reviewer respects the repository's SECURITY.md.

Adds explicit security-review output and repo-specific security policy awareness — directly relevant for teams using Devin in security-sensitive codebases.

MCP / tool calling / API

June Cline CLI releases expanded the agent/plugin/MCP surface: plugin commands can submit prompts to the agent, plugins gained MCP server support with OAuth during install, Cline added `cline skill`, and the CLI added a prefilled MCP install wizard. The same releases added output caps and ranged reads for large files.

Materially improves tool/plugin extensibility, MCP server setup, and safer handling of large tool outputs — relevant for teams building on or extending Cline.

MCP / tool calling / API

Activepieces release 0.85.4 added AI-ready metadata across hundreds of pieces, tagged non-agent actions with audience='human', and updated builder docs to require AI metadata on new actions and triggers.

Improves tool discoverability, tool suitability labeling, and the quality of AI-facing action and trigger metadata — relevant for teams integrating Activepieces pieces into agentic workflows.

Observability / auditability

Activepieces release 0.85.4 shipped broader runtime visibility improvements, including evlog-wide events replacing OpenTelemetry + pino, date ranges for queue health metrics, and worker/sandbox state and RAM visibility.

Strengthens operator visibility into worker, sandbox, event, and queue health — relevant for teams running Activepieces in production environments that require operational observability.

Workflow orchestration

Activepieces release 0.85.4 added error handlers to steps in flows.

Improves failure handling and recovery behavior inside multi-step automations — relevant for teams building resilient workflows where individual step failures should not abort the entire flow.

MCP / tool calling / API

langgraph-cli 0.4.30 added support for compatible API version ranges.

Version-range compatibility reduces operational friction between CLI tooling and server/API deployments.

Observability / auditability

Agent Chat Evaluations let teams define criteria and automatically grade agent chats.

Adds native QA/evaluation instrumentation for production agents instead of relying only on external eval tooling.

Human approval / guardrails

Human-in-the-loop for agents lets agents pause mid-task to request approval before running a tool, or ask a question with options, then resume where they left off in agent chats and Slack.

A meaningful governance/control upgrade for agent deployments where autonomous tool use needs checkpoints.

Integrations

Clari Copilot integration became generally available to all Dust workspaces via MCP server configuration.

Adds enterprise sales-call intelligence as an agent-usable source/action surface.

Memory / state

Email replies now stay in a single Dust conversation instead of creating a new conversation for each thread reply.

Material state and continuity improvement for email agents — preserves conversation history across thread replies.

Funding / partnership

Bland announced an additional $50M in funding, bringing total raised to more than $100M. Named backers include Scale, Emergence, HubSpot, Dell Technologies Capital, and Upfront.

A material company-level signal for durability, hiring capacity, and enterprise go-to-market momentum.

Deployment / data residency

Tray Headless MCP added region-specific endpoints for US, EU, and APAC workspaces.

Regional MCP endpoints affect deployment geography and handling for MCP-compatible clients.

Memory / state

Tool responses can now be saved to properties, not just variables.

Expands how tool output can persist in project state and metadata — relevant for downstream tagging, tracking, and memory-like state handling.

MCP / tool calling / API

Retell removed legacy list endpoints and pushed users to versioned v2/v3 list APIs with unified pagination.

Material API compatibility and integration-maturity change affecting list/query behavior for teams building on Retell's API.

Funding / partnership

Salesforce signed a definitive agreement to acquire Fin for about $3.6B, with closing expected in Salesforce fiscal Q4 2027.

Major ecosystem and go-to-market event with likely downstream impact on enterprise reach, partner strategy, and roadmap signaling.

Workflow orchestration

ElevenLabs JS and Python SDK v2.53.0 added conversation evaluation rerun support, workflow entry_behavior and end-procedure tools, telephony filters, and expanded agent test batch limits.

Material for SDK-driven testing, evaluation, and workflow controls in voice agents.

MCP / tool calling / API

Dust became connectable as a remote MCP server to any MCP-capable client. Admins can disable MCP access and restrict allowed redirect URIs.

Major interoperability change — Dust agents can now be invoked from any MCP-compatible client, with meaningful admin-control implications for enterprise deployments.

Funding / partnership

Decagon announced an accredited integration with Five9 and availability on the Five9 CX Marketplace.

Material enterprise ecosystem and distribution change for contact-center deployment — relevant for buyers evaluating Decagon in CCaaS environments.

Security / enterprise

Custom Roles API added an `org_id` parameter for enterprise customers using workspace-scoped API tokens.

Improves enterprise API governance and organization-specific access control — relevant for large teams managing multiple repositories or workspaces.

Agent capability

CodeRabbit now runs oasdiff during PR reviews to detect breaking changes between OpenAPI specifications.

Adds API-aware review capability for detecting breaking contract changes — relevant for teams with public or partner-facing APIs.

Human approval / guardrails

Claude Code v2.1.178 added parameter-level permission rules using Tool(param:value) syntax and changed auto mode so sub-agent spawns are classifier-checked before launch.

This strengthens tool governance and sub-agent launch controls — relevant for teams that need fine-grained control over what tools agents can invoke and how sub-agents are authorized before execution.

MCP / tool calling / API

Agno v2.6.15 made the AgentOS MCP server a configurable extension point via MCPServerConfig, with custom tool registration, built-in tool scoping, caller-identity injection, authorize hooks, and allowed_hosts / allowed_origins protection.

A major MCP and tool-governance update. It expands agent interoperability while adding stronger controls around tool exposure and caller identity — relevant for enterprise teams requiring scoped, auditable MCP integrations.

Workflow orchestration

Databricks announced Omnigent, an open-source meta-harness above existing agents that adds multi-agent composition, contextual policies, real-time collaboration, cloud execution, and multi-harness authoring.

Direct index-relevant shift toward cross-agent coordination, policy control, collaboration, and agent operations rather than a single-agent capability update.

Deployment / data residency

deepagents-code 0.1.16 added a Vercel Sandbox provider and ChatGPT OAuth sign-in for Codex models.

Expands execution-environment options for agent code and reduces authentication friction for Codex-based workflows.

Agent capability

Workspace-level Agent customization added reusable workspace instructions plus on-demand skills.

Changes how teams standardize agent behavior across projects, improving consistency and repeatability.

Security / enterprise

Package Firewall now blocks malicious or compromised packages by default, and Agent adapts its plans to safer alternatives.

Material governance and security improvement for production coding agents — reduces supply-chain risk in agent-assisted builds.

Memory / state

LangGraph 1.2.5 fixed DeltaChannel state updates on empty threads and merged lc_versions config metadata.

These fixes affect correctness of persisted state and version metadata in long-running, stateful agent graphs.

Security / enterprise

Devin added custom OAuth configuration for Marketplace MCP servers and an enterprise setting to enable or disable web search.

Strengthens enterprise control over MCP authentication and external web-search behavior — key governance controls for regulated or security-conscious deployments.

Deployment / data residency

Braintrust released terraform-aws-braintrust-data-plane v5.5.0, updating Braintrust Services/API ECS to v2.2.1, changing the default Redis instance type from cache.t4g.medium to cache.r7g.large, and enabling Brainstore fast readers by default.

Materially affects self-hosted AWS deployment defaults, capacity assumptions, and upgrade behavior for teams running Braintrust on their own AWS infrastructure.

Observability / auditability

Claude Code v2.1.174 added usage attribution in /usage showing cache misses, long context, subagents, and per-skill/agent/plugin/MCP breakdowns over 24h and 7d windows in VS Code.

Improves visibility into where agent usage is coming from across tools, plugins, subagents, and MCP — relevant for teams managing cost attribution and operational oversight in multi-agent deployments.

Memory / state

Agno v2.6.14 added create, read, update, and delete endpoints for AgentOS learnings.

CRUD access to learnings is a concrete memory and state management capability for agent persistence and operational memory — relevant for teams building agents that need to accumulate and manage knowledge over time.

MCP / tool calling / API

Agent Bricks Supervisor Agent added support for Unity Catalog volumes as subagent tools.

Expands the tool surface available to supervised multi-agent systems and is a concrete tool-access orchestration update.

MCP / tool calling / API

Outreach released Outreach MCP Client for Omni, letting reps pull knowledge and actions from third-party MCP connectors, browse/install marketplace connectors, and use private MCP connectors restricted to their instance.

Adds MCP consumption, marketplace connectors, and private connector controls inside the agent surface — a clear MCP/index-relevant capability change.

Memory / state

Outreach added Knowledge Enhancements surfaced in Omni, allowing admins to upload approved documents, index them, assign channels, validate them in Playground, and expose cited source documents in Omni responses.

Materially changes Omni's grounding layer: enterprise memory becomes admin-curated, channel-scoped, testable, and source-visible.

Browser/computer use

Outreach added Web Search in Personalization Agent, enabling real-time external web context, prompt-controlled retrieval behavior, and visible source citations in preview.

Direct web-augmented agentic capability: externally grounded personalization with reviewable provenance before send.

Workflow orchestration

Outreach consolidated Smart Account/Deal Assist into Omni on Account, Prospect, and Opportunity, routing account and deal queries through Omni across multiple entry points.

Product-surface orchestration change: the AI experience becomes centralized rather than split across separate assistants.

Observability / auditability

Outreach released Individual Revenue Agent Reporting, adding direct performance-report links from Revenue Agent list and overview views to agent-scoped Sequence Performance Reports.

Strengthens measurement of autonomous/semi-autonomous revenue-agent outcomes with agent-level performance reporting.

Security / enterprise

langgraph-cli 0.4.29 added support for passing a certfile and cert key to run the dev server under HTTPS.

Improves secure local/staging deployment ergonomics for agent apps and agent-server testing.

Integrations

CrewAI 1.14.7 added native Snowflake Cortex LLM provider support and Databricks / Snowflake integration guides.

Expands enterprise data and AI platform integration surface for teams running CrewAI alongside Snowflake or Databricks.

Observability / auditability

CrewAI 1.14.7 surfaced real `finish_reason`, sampling parameters, and `response.id` on LLM events.

Improves observability into LLM calls and execution outcomes — useful for teams debugging agent behavior or auditing model usage.

Workflow orchestration

CrewAI 1.14.7 added chat API for conversational flows, route-aware DSL triggers, and FlowDefinition construction from Flow DSL metadata.

Expands CrewAI's workflow orchestration and conversational flow surface, enabling more dynamic agent routing and flow construction.

Memory / state

CrewAI 1.14.7 added pluggable default backends for memory, knowledge, RAG, and flow.

Materially improves CrewAI's configurability around memory/state and orchestration infrastructure — teams can now swap storage/retrieval backends without forking.

Integrations

Finishing Touches expanded platform coverage: Autofix on Azure DevOps PRs, plus Generate unit tests and Custom recipes on GitLab merge requests.

Broadens CodeRabbit's agentic review and remediation capabilities across additional code-hosting surfaces.

Memory / state

CodeRabbit added Learning approvals so admins can review newly created learnings before they enter the knowledge base.

Adds governance over agent memory and knowledge-base mutation — relevant for teams where incorrect learnings could affect review quality.

Security / enterprise

CodeRabbit added SkillSpector scanning for AI agent skills and MCP configuration files during PR reviews, inspecting SKILL.md, MCP configs, Claude desktop config, Cursor rules, and Codex YAML for risks.

Directly relevant to agent security — provides automated risk detection for AI agent skill definitions and MCP configurations in code reviews.

Integrations

CodeRabbit IDE extension added direct Devin handoff links.

Strengthens cross-agent workflow support by allowing CodeRabbit findings or plans to move directly into Devin for implementation.

Memory / state

Braintrust released bt CLI v0.12.0, adding dataset snapshot create/list/restore/delete support plus bt topics btmap download.

Dataset snapshots improve reproducibility and state/version control for eval datasets. Topic-map artifact export improves analysis workflows.

Security / enterprise

Sierra announced FedRAMP High certification, making its agent platform available for U.S. federal-agency deployment.

FedRAMP High materially changes Sierra's enterprise/government eligibility and security/compliance posture for regulated deployments.

Memory / state

Voiceflow added Personas, letting teams test agents with predefined starting variable values for specific customer types or account states.

Adds reusable stateful testing context and improves repeatability for agent evaluation.

MCP / tool calling / API

Kapa launched Kapa for Agent Context, a retrieval layer for agents available via Hosted MCP Server or Retrieval API, grounded in docs, code, PDFs, tickets, Slack, and 30+ sources, with citation-backed fallback behavior.

Clear shift from 'AI assistant for documentation' toward agent-context infrastructure and MCP-native grounding — relevant for teams building agents that need authoritative, cited knowledge retrieval.

Human approval / guardrails

Fin over Email added pre-live reply preview/testing, a dedicated Spam view with customizable spam guidance, and channel-specific guidance/escalation controls.

Meaningful safe-deployment controls — preview/testing before go-live, spam governance, and per-channel behavior separation are relevant for teams deploying Fin in regulated or high-volume email environments.

Workflow orchestration

Fin over Email now sends configurable follow-up emails to customers who go quiet, with delays from 1 hour to 30 days, across Email, Zendesk, Salesforce, Freshdesk, and HubSpot.

Changes case-closure and re-engagement automation behavior in agent-driven support workflows.

Funding / partnership

E2B joined the Stripe Projects developer preview so agents can discover, provision, and authenticate E2B sandboxes without manual API-key setup.

Material ecosystem change for agent runtime provisioning and credential flow — lowers the integration barrier for teams building agentic systems with E2B sandboxes.

Observability / auditability

Bugbot became roughly 3x faster, about 22% cheaper, and found about 10% more bugs per review. /review can now run Bugbot and Security Review before push, with diff deduplication across local reviews and GitHub/GitLab PRs.

Material change to agentic code-review performance, cost, and pre-push workflow coverage — relevant for teams using Cursor's automated review capabilities.

Observability / auditability

Devin Review now cancels pending PR reviews when new commits arrive, and the PR Review API exposes a new cancelled status.

Improves auditability and state tracking for automated PR reviews — ensures stale reviews don't complete on outdated code.

MCP / tool calling / API

Devin added the official Figma MCP integration and deactivated the previous unofficial integration.

Adds an official MCP-based design-tool integration, replacing an unofficial path — relevant for teams using Devin in design-to-code workflows.

Workflow orchestration

CodeRabbit Plan became available directly in the VS Code extension, with a Plans tab for creating agent-ready Coding Plans and handing phases to configured AI coding agents.

Moves planning/orchestration into the IDE and makes CodeRabbit more directly useful in agentic coding workflows.

Workflow orchestration

Cline CLI v3.0.23 added support for configured agents as subagent tools and centralized OAuth management into the SDK.

A concrete subagent/tool orchestration expansion — configured agents can now be invoked as tools, improving multi-agent workflow composition.

MCP / tool calling / API

browser-use released 0.13.1, adding support for Claude Fable 5 and changing Anthropic behavior so tool choice is auto when using thinking. This is a dated compatibility release; current Fable 5 availability has not been separately verified.

Post-cutoff compatibility change relevant to model support and tool-calling behavior for browser-use deployments using Anthropic models. Buyers should verify current Fable 5 availability directly.

Workflow orchestration

Claude Code v2.1.172 enabled sub-agents to spawn their own sub-agents, up to five levels deep.

This materially expands recursive and delegated agent orchestration depth, relevant for teams building complex multi-agent coding pipelines or automated software-engineering workflows.

Human approval / guardrails

Agno v2.6.13 added socket support for human-in-the-loop workflows.

Directly relevant to approval loops and interactive workflow orchestration — socket-based HITL enables real-time human intervention in running agent workflows.

Integrations

Tray.ai updated its OpenAI connector to improve token management and model-selection logic across model families.

Real integration update for agent builders using Tray's OpenAI connector with model-family-aware token handling.

Observability / auditability

Voiceflow introduced Tests, which simulate conversations and check whether responses, routing, and tool calls match expected behavior before production exposure.

Material verification and QA capability for agent builders, directly affecting how reliably agent behavior can be validated before go-live.

Observability / auditability

Salesloft added three AI-usage metrics in Analytics: Account researched, Person researched, and Agent tasks completed.

Relevant to measurable agent adoption and observability — teams can now track AI task completion alongside traditional activity metrics.

MCP / tool calling / API

Salesloft added direct connection to Claude through the MCP Connectors Directory for customers with the Agentic add-on.

Expands the agentic surface area beyond the native UI — clear MCP/connectors change gated by the Agentic add-on.

Workflow orchestration

Cadence Collections introduced a new parent container for grouping cadences and inheriting settings, permissions, and CRM sync rules.

Changes how teams govern and orchestrate repeatable engagement workflows at scale.

MCP / tool calling / API

Agentforce Vibes 2.0 added an Abilities & Skills framework that dynamically activates domain-relevant context including MCP tools, plus additional frontier models and React app generation.

Changes how Agentforce-based builders access tool/context scaffolding and expands supported model and UI outputs.

Workflow orchestration

NICE announced that agentic AI is native at the core of the CX platform, naming NICE AI Agents, the Agentic Engagement Plane, Guardian AI, and Agentic Analytics as key layers of the operating model.

Formalizes orchestration, governance, and analytics layers as first-class platform components rather than ancillary messaging.

Observability / auditability

NICE launched the CXone Workforce Empowerment Suite for managing human and AI agents under one operating model, with shared dashboards, quality/compliance controls, and AI operations visibility.

Relevant as hybrid human/AI workforce governance and monitoring, especially for enterprise CX operations.

Observability / auditability

LlamaParse added opt-in granular bounding boxes at line, word, and cell level, with beta availability across paid tiers.

Granular provenance materially improves auditability, citation precision, and human verification in document-grounded agent workflows.

Integrations

Fin can run as a service agent on top of HubSpot and Freshdesk without requiring migration off the current helpdesk.

Major platform-openness and deployment-surface expansion — buyers on HubSpot or Freshdesk can now deploy Fin without switching helpdesks.

MCP / tool calling / API

Haystack 2.30.1 lets AzureOpenAIChatGenerator accept Secret values for azure_endpoint and api_version, enabling environment-variable-based runtime switching.

Production-configuration improvement for teams moving the same serialized pipeline across environments without hardcoding credentials.

MCP / tool calling / API

Large MCP toolsets now lazy-load tool details only when needed, improving context efficiency and reducing unnecessary token usage.

Materially improves tool-heavy agent scalability and lowers context/cost pressure in large MCP deployments.

Integrations

Agents can now be connected to Microsoft Teams as an external channel.

Broadens deployment surfaces into a major enterprise collaboration environment.

Agent capability

Decagon announced Duet Autopilot, which turns production signals into validated agent updates ready for human review. Decagon also announced DuetBench alongside it.

Material change toward self-improving agents with explicit human-review gating — relevant for buyers evaluating autonomous quality improvement in customer-support agents.

Security / enterprise

Composio CLI 0.2.31 included security dependency updates: authlib bumped to 1.7.2 for GHSA-wvwj-cvrp-7pv5 and protobufjs pinned to 7.5.5 for a critical Socket.dev CVE.

Security maintenance relevant for developer tooling and API/MCP infrastructure. Low-impact public index update but important for security-conscious teams.

Integrations

Automatic Repository Linking can now discover and link related repositories across an organization so cross-repo context appears in reviews.

Improves cross-repo context and repository-level memory for code review workflows in multi-repo organizations.

Agent capability

Cline v3.89.0 added Claude Fable 5 model support. This is a dated release event; current Fable 5 availability has not been separately verified.

Model availability affects the coding agent's supported backend capability set — buyers should verify current Fable 5 availability directly with Cline/Anthropic.

Security / enterprise

Stagehand server-v3 v3.7.2 added Azure Entra model authentication support.

Improves enterprise authentication posture for Browserbase / Stagehand model-backed browser-agent infrastructure — relevant for enterprise teams requiring Entra-based auth.

Pricing / packaging

API access through API keys became available on Beam's Free plan.

Changes programmatic-access gating and lowers the threshold for experimentation, evaluation, and integration.

Human approval / guardrails

Beam added "Test Before You Publish," a pre-publish testing mode that runs flows against historical real data in a safe environment, sandboxes integrations, retains test execution review, and warns if live connections remain active before publish.

A meaningful governance and control upgrade for agentic workflows, especially when integrations can mutate external systems.

Agent capability

Augment Code added Claude Fable 5 to the model picker, positioning it for long, multi-step, deep-reasoning work. This is a dated release event; current Fable 5 availability has not been separately verified.

Expands the model surface available to Augment's agent workflows — buyers should verify current Fable 5 availability directly with Augment.

Agent capability

Claude Code v2.1.170 added access to Claude Fable 5 as of this release date. This is a dated release event; current Fable 5 availability has not been separately verified.

Model/backend availability is index-relevant when it materially changes the agent's available reasoning or coding capability — buyers should verify current availability directly with Anthropic.

Workflow orchestration

Activepieces release 0.85.2 added data manipulation triggers.

Expands trigger-layer flow input and transformation capability, which is relevant to workflow flexibility and pre-processing in agentic automation pipelines.

Agent capability

Salesforce added Google Gemini as a selectable model for Agentforce / the Atlas reasoning engine.

Material model-choice expansion for Agentforce deployments — buyers can now select Gemini alongside existing model options.

Workflow orchestration

Business Details headers became configurable with up to 12 fields and org-wide/default layouts.

Changes investigator workflow design and layout governance for business-review operations.

Observability / auditability

Device Intelligence now supports up to three custom charts, and the prior map visualization was removed.

Improves risk-team monitoring surfaces with configurable dashboard visualizations.

MCP / tool calling / API

Sardine added multi-party AML screening for wire transactions so one API call screens sender, sender bank, receiver, receiver bank, and memo.

Material API/screening capability expansion — reduces integration overhead for compliance teams requiring multi-party wire coverage.

Observability / auditability

Screening widgets now support batch entity resolutions across sanctions, PEP, and adverse media lists.

Reduces manual queue handling and improves audit consistency in high-volume review operations.

MCP / tool calling / API

ElevenLabs changed the default conversational ASR provider from elevenlabs to scribe_realtime and added conversation_product to conversation details responses.

Concrete API/runtime changes for conversational-agent speech processing and metadata — relevant for teams building on or integrating ElevenLabs Conversational AI.

Integrations

Gamma became available as a Remote MCP Server in Dust for generating and reading Gamma decks, docs, and webpages.

Adds a new MCP-based artifact generation and output surface for teams using Dust in content-creation workflows.

Workflow orchestration

Dust introduced skill composition, letting skills reference other skills and tools inline in instructions. The separate tools section was removed.

Material orchestration change for reusable multi-step capabilities — enables modular, composable agent skill design.

Agent capability

Builders can ask @dust to create or edit skills directly from a conversation.

Materially lowers friction for agent-skill authoring and iteration — no need to leave the conversation to update agent capabilities.

Browser/computer use

browser-use released 0.13.0, "Rebuilt in Rust [beta]," introducing a new Rust-backed beta agent with a more direct browser-control loop. The existing Python agent remained unchanged.

A major browser-agent architecture change — the Rust-backed beta agent directly affects how browser/computer-use tasks are executed.

Agent capability

CrewAI published 1.14.7a1 (June 3) and 1.14.7a2 (June 5) with trained-agents file support, native Snowflake Cortex LLM provider, Databricks and Snowflake integration guides, conversational flow traces, a chat API for conversational flows, route-aware flow DSL work, more detailed LLM event surfacing, and flow/lock-store refactoring.

Meaningful OSS movement for CrewAI's agent orchestration and conversational-flow direction. These are pre-release/alpha builds, not stable GA releases — buyers should evaluate accordingly before relying on these capabilities in production.

Pricing / packaging

Cognition introduced an AI Productivity Guarantee for eligible enterprise Devin customers: if Devin delivers less engineering value than the customer pays for, Cognition will issue usage credits up to $10M. Value is measured in equivalent engineering hours using an estimator that reviews completed Devin sessions.

This is a rare commercial guarantee tied to measurable agent output, and it materially affects enterprise buyer conversations around ROI, accountability, and spend governance.

Agent capability

Activepieces v0.85.0 added Mistral AI as a platform AI provider, formula/data-manipulation functions in the flow builder, platform-admin visibility into run success rates/live queue depth/30-day health history, clearer failed-step errors with one-click AI help, AI chat reliability improvements, model tiering with native Anthropic thinking, improved tool-calling UX, pure tool-calling parsing, multi-item action execution with live progress, display tools with agentic rendering, and AI badges for MCP-created flows.

A meaningful platform release touching agent execution, AI chat, model/provider support, workflow building, MCP labeling, and operations visibility.

Workflow orchestration

Haystack v2.30.0 added PythonCodeSplitter for syntax-aware Python code splitting in code-RAG/code-search pipelines; allowed ChatGenerator components to accept plain strings as messages; updated DALL-E image generation defaults for OpenAI's newer image models; added async retriever support; and fixed an Agent bug where multiple tool calls could prevent the configured exit condition from stopping the loop.

The most material framework-level update in this review window: touches code-centric agent/RAG workflows, developer ergonomics, async execution, model migration, and agent control-flow reliability.

Agent capability

Augment announced Cosmos, a platform for AI-native engineering teams. Cosmos is positioned around an agentic SDLC where agents operate across triage, spec, implementation, review, testing, deployment, and feedback; it supports teams of specialized agents that coordinate, delegate, and share memory. Available to all plans and every team plan.

This shifts Augment's positioning from coding assistance toward a broader software-engineering agent platform, with workflow orchestration, shared memory, MCP/webhook integrations, and multi-agent coordination.

MCP / tool calling / API

Gumloop 9.11.0 expanded the Gumloop MCP server with tools for agents, sessions, skills, files, teams, connected MCP servers, resources, prompts, and MCP tool calls. The release also added Looker MCP support and BigQuery Workload Identity Federation.

Improves Gumloop's value as an agent/workflow platform that can expose and govern more tools, enterprise data systems, and MCP-accessible capabilities.

Security / enterprise

Gumloop 9.11.0 added the ability to add skills from Shared With Me and Organization tabs directly to agents, and introduced skill permission roles — Editor, Viewer, and Use Only — so teams can govern who can update, inspect, or use skills inside agents.

An enterprise-relevant governance update for reusable agent skills, permissions, and multi-user workspace control.

Agent capability

Langflow 1.9.6 fixed Agent LLM streaming so tokens stream incrementally, and included a remote component error-handling fix for Docling.

Improves runtime behavior and user experience for agent execution and streaming responses.

Agent capability

Cognition announced Devin Desktop, repositioning Windsurf around the Agent Command Center as the default IDE surface for managing local and cloud agents, PRs, and context. The release adds Spaces so related agents can share context, and adds Agent Client Protocol (ACP) support so ACP-compatible agents can run inside Devin Desktop alongside Devin.

A major product-surface shift from IDE/coding assistant toward multi-agent management for software engineering teams.

MCP / tool calling / API

Langflow 1.9.5 shipped a fix for model handling for tool calling in agents and updated IBM models. The release also included documentation updates around AGENTS.md and project philosophy docs.

Incremental but material for framework users relying on agent tool-calling and IBM model support.

Integrations

Activepieces released v0.84.0 with new integrations including Resend, Plausible Analytics, PostHog, UptimeRobot, Azure DevOps, Frill, Beebole, YouTrack, Hootsuite, PDF4me, Sendr, and Streak CRM, plus GitHub App authentication and safer handling for disabled subflows.

A broad integration push across dev-ops, analytics, CRM, and productivity tools extends Activepieces' automation reach without custom connectors.

Agent capability

Devin added Auto-Triage for autonomous issue prioritization, native Windows VM support for cross-platform development tasks, and broader deployment and financing announcements signaling expanded enterprise ambition.

Auto-Triage and Windows VM support expand Devin's autonomous engineering footprint into issue management and cross-platform builds — material for enterprise teams evaluating end-to-end coding agent coverage.

Human approval / guardrails

Cursor shipped Composer 2.5, Jira integration, shared canvases, /loop, and Auto-review Run Mode for safer long-running Shell, MCP, and Fetch execution — using allowlists, sandboxing, and classifier-based approval logic to control what agents can run autonomously.

Auto-review Run Mode with classifier-based approval is a significant guardrails addition for teams concerned about autonomous agent actions in production codebases and CI pipelines.

MCP / tool calling / API

Gumloop's late-May releases added MCP Artifacts, queued mid-run steering messages, hosted agent pages, richer HTML artifacts, Freshdesk and Freshsales MCP, Apify MCP, team-level secrets, a notification center, and Claude Opus 4.8 support.

MCP Artifacts and mid-run steering messages strengthen Gumloop's case for complex, interactive agentic workflows where runtime human guidance and structured tool outputs matter.

Security / enterprise

CrewAI released v1.14.6 with StdioTransport protections against environment-variable leakage, improved planning and observation handling, and fixes for structured output leakage, checkpoint restore, and executor resume behavior.

Environment-variable and structured-output leakage protections directly address security concerns for teams running CrewAI agents with sensitive credentials or proprietary data.

MCP / tool calling / API

n8n shipped reliability fixes for binary-data temp-directory cleanup and webhook close-function draining to prevent MCP connection leaks.

MCP connection leak prevention is directly relevant for production deployments where long-running n8n agent workflows depend on stable MCP server connections.

Funding / partnership

Cognition announced it raised over $1B at a $26B valuation, led by Lux Capital, General Catalyst, and 8VC. Enterprise usage is up more than 10x since the start of the year and run-rate revenue is at $492M.

Materially changes the vendor's market position, enterprise traction profile, and funding/scale narrative — relevant context for any enterprise evaluation or vendor-longevity assessment.

Observability / auditability

Mastra added stored-entity HTTP APIs, browser session probing, screencast support, delta-polling observability, stricter authorization defaults, Convex native vector search, Google Cloud Spanner storage, upgraded agent channels, client-side tool observability, and a FilesSDK-backed workspace filesystem provider.

Delta-polling observability, client-side tool visibility, and stricter auth defaults are enterprise-readiness signals for JS/TS teams evaluating Mastra for production agent deployments.

Workflow orchestration

Augment Code's Cosmos releases added service-account attribution, scoped webhooks, hosted artifacts, broader Slack controls, file attachments in web conversations, and simplified custom environment requirements (bash and git only).

Service-account attribution and scoped webhooks improve traceability and governance for enterprise teams running Augment in CI/CD and collaborative coding workflows.

Browser/computer use

browser-use shipped 0.12.7–0.12.9 with a major CLI update, tighter daemon socket permissions, restricted browser-profile handling, cached-content and prompt-history improvements, and new-tab screenshot reliability fixes.

Tighter socket permissions and restricted profile handling address security hygiene for teams running browser-use in shared or production environments.

Agent capability

Agno added support for Google Antigravity and Gemini managed agents for Deep Research, plus approval metadata in post-hooks, PgVector prefix-match fixes, and a Gemini tool-call path correction.

Native Gemini Deep Research and Antigravity integration expands Agno's model-agnostic positioning and strengthens its case for teams building research-grade multi-modal agents.

Browser/computer use

Browserbase updated its Fetch API to return markdown and structured JSON, raised the response cap from 1 MB to 5 MB, and positioned Fetch as a lower-cost read path than launching a full browser session.

A structured-output Fetch API at lower cost than full browser sessions meaningfully reduces cost and latency for agents doing web data extraction without needing full browser automation.

Security / enterprise

Dify released v1.14.2 with security hardening, workflow and HITL reliability fixes, RAG and document-processing improvements, knowledge-base stability updates, tracing reliability, and self-hosted deployment and runtime changes.

Continued hardening across security, HITL, and knowledge-base stability is relevant for self-hosted production deployments where reliability and data governance are evaluation criteria.

Security / enterprise

Sourcegraph added granular admin permissions via RBAC, allowing teams to delegate specific admin capabilities without granting full administrator access.

Granular RBAC is a foundational enterprise governance requirement — directly relevant for large teams and regulated industries where least-privilege access is a procurement criterion.

Pricing / packaging

MindStudio sharpened its agent-pricing position around token-based usage, free and Individual tiers, 200+ model access, and Business-tier controls such as SSO, audit logs, usage alerts, and unlimited collaborators.

Useful for buyers comparing agent platforms against SaaS-native "second meter" pricing, lock-in risk, and enterprise governance requirements.

Workflow orchestration

Gumloop's 9.5.0 changelog added Shared With Me and Organization views across agents, skills, files, and workflows, plus consolidated agent activity in the Home tab.

Helps teams move from individual automation projects to shared departmental workflows with better discoverability, handoff, and operational visibility.

Workflow orchestration

Lyzr published a finance-ops case study on payment-receipt verification automation, including portal, email, and WhatsApp intake, payment-method guardrails, API connectivity, and manual exception routing.

Useful customer evidence for buyers evaluating agentic automation in finance operations, especially where exception handling and system integrations matter.

Funding / partnership

11x published a post claiming Ramp spend data ranked it first in enterprise adoption for AI-first GTM, with 40 percent of enterprises buying in that category choosing 11x.

Useful as a market-validation signal, but should be treated as a vendor-stated procurement-data claim rather than a fully transparent public benchmark.

Browser/computer use

Cognition added Android emulator support for Devin, allowing it to run Android Virtual Devices, build and inspect Android apps, reproduce bugs, and verify changes locally.

Expands Devin's autonomous verification surface for mobile engineering teams and improves its credibility for Android testing, debugging, and app-development workflows.

Agent capability

Decagon introduced Guided Discovery, a capability for exploratory product-discovery, retention, and expansion conversations without relying entirely on predefined scripts.

Signals Decagon is expanding beyond rigid support-ticket automation into concierge-style customer conversations, commerce, retention, and expansion use cases.

Security / enterprise

Dify released v1.14.1, a security and stability patch covering self-hosted SECRET_KEY hardening, internal metrics protection, an IDOR fix, tenant-scoped credential cleanup, LiteLLM CVE remediation, and workflow, HITL, and knowledge-base fixes.

Meaningful for self-hosted and production users because it addresses concrete security, reliability, workflow, and knowledge-base risks.

API / Governance

Devin added a Review API for programmatic pull-request review, per-PR auto-review toggles, service-user permission management, and MCP secret scoping.

These updates strengthen Devin's enterprise story around CI/CD integration, review automation, auditability, permissioning, and secure engineering-agent operations.

Decagon's resources page added a customer clip showing how GlossGenius uses Duet to scale concierge customer experiences.

This is useful buyer evidence for Decagon's Duet adoption story, but should be treated as customer validation rather than a net-new platform capability.

Feature / Orchestration

Gumloop announced Subagents, allowing agents to clone themselves for parallel work and call other specialized agents.

This strengthens Gumloop's position for buyers comparing multi-agent orchestration, delegated workflow execution, and more complex agentic automation patterns.

Product / Positioning

CrewAI published Discovery, a new engine designed to surface the best automation use cases for a business. May 4–8 open-source releases included LLM listing refreshes, a status endpoint fix, gitpython security compliance, and CLI packaging cleanup.

This moves CrewAI's positioning upstream from agent orchestration into automation prioritization, which matters for buyers still trying to identify where agentic workflows are most likely to produce ROI.

Integration / Governance

Gumloop's 9.2.0 changelog added Gmail triggers for any label, audit-log filters, Google Analytics MCP, and Outlook Calendar MCP.

The update broadens Gumloop's operational integration surface while improving audit/log filtering for teams that need better workflow visibility and governance.

Agent capability

Cursor launched Background Agents, enabling long-running autonomous coding tasks that execute in the cloud without requiring the IDE to remain open — agents can clone repos, run code, and open PRs independently.

Background Agents represent a meaningful shift from copilot-style assistance to fully autonomous software engineering, directly relevant for teams evaluating agentic coding platforms.

Integrations

Relevance AI added Confluence Knowledge Sync, project-level public-agent controls, and an Android app, improving enterprise knowledge connectivity, governance, and mobile workforce access.

Relevant for buyers evaluating RAG/knowledge grounding, governance over public-agent sharing, and operational access for mobile-first teams.

Funding / partnership

Decagon announced a new APAC office to support increasing regional demand for its AI concierge platform and enterprise customers in the region.

Signals commercial maturity, expanded customer support footprint, and growing regional enterprise traction — relevant for APAC buyers evaluating vendor longevity.

Workflow orchestration

Gumloop's late-April releases added subagents, agent email inboxes, an organization analytics agent, MCP hosting/proxying, enterprise data drains to S3 and BigQuery, Google Search Console MCP, and higher Pro credit limits.

A major release cycle across multi-agent coordination, observability, MCP infrastructure, enterprise export, and packaging — relevant for buyers evaluating Gumloop for production-scale deployments.

Human approval / guardrails

Dify v1.14.0 added real-time collaborative workflow editing, a service API for human-in-the-loop, multiple MCP and OAuth fixes, optional Langfuse TTFT reporting, billing/quota improvements, stronger tenant checks in knowledge APIs, and security hardening.

A substantial release for self-hosted and enterprise evaluators covering collaboration, HITL, MCP reliability, observability, billing, data governance, and security.

Integrations

StackAI introduced a Box integration for knowledge bases, workflows, conversational assistants, triggers, and write-back to Box. Also published a deployment-options guide covering multi-tenant cloud, dedicated cloud, private cloud/BYOC, and on-premise.

Improves enterprise content integration and gives buyers clearer deployment, residency, and compliance context when evaluating StackAI.

Workflow orchestration

Beam AI added a Code Execution Node that lets builders run JavaScript or Python directly inside flows without LLM overhead. Also added zip-file trigger support and Outlook attachment public-URL handling.

Expands Beam from LLM orchestration toward deterministic workflow execution inside agentic applications — relevant for buyers needing reliable, cost-efficient compute steps.

Security / enterprise

Cognition launched Devin for Terminal, enabling local-to-cloud session handoff, alongside new enterprise controls including network policy restrictions and more granular PR review permissions.

Improves local-to-cloud coding-agent workflows and strengthens enterprise governance and security evaluation criteria.

Observability / auditability

Decagon introduced automatic optimization and Root Cause Analysis through Duet.

Automated quality improvement and RCA tooling reduces manual QA overhead in production support deployments.

Human approval / guardrails

Gumstack added App Policies for enterprise control over how agents use apps and data.

App Policies give enterprise buyers programmatic guardrails over agent behaviour, a key requirement for regulated-industry deployments.

Agent capability

Ava 2.0 is publicly marketed by Artisan, but official pages conflict on whether it is fully live or still launching. This item requires manual review before relying on capability claims.

If confirmed live, Ava 2.0 represents a significant upgrade to Artisan's outbound sales agent. Buyers should verify directly with the vendor before making decisions based on this update.

Workflow orchestration

LangGraph Platform reached general availability, providing a managed cloud runtime for deploying, scaling, and monitoring stateful multi-agent LangGraph workflows, with support for persistence, streaming, and human-in-the-loop interrupts.

GA status means production-grade SLAs and managed infrastructure for teams that want LangGraph's orchestration model without self-hosting the backend.

Human approval / guardrails

CrewAI changelog and docs verify Flow human-in-the-loop (HITL), user input handling, and HITL loop behaviour across recent releases.

Native HITL in CrewAI Flows enables supervised agentic workflows without external orchestration tooling.

Workflow orchestration

Devin can now manage parallel child Devins and schedule recurring sessions with persistent state across runs.

Multi-agent orchestration with scheduling and state persistence is a critical capability signal for teams evaluating autonomous coding workflows at scale.

Funding / partnership

Gumloop raised $50M Series B and positioned Gumstack as a security and observability layer for organizational agents.

Series B funding and a security-positioning shift signals enterprise ambition and product longevity.

Agent capability

Decagon introduced proactive agents, user memory, and outbound campaigns.

Proactive outreach and persistent user memory represent a significant leap beyond reactive support agents, relevant for teams evaluating autonomous customer engagement.

Deployment / data residency

AGI Inc. announced a Qualcomm collaboration to bring agentic AI to Snapdragon-powered devices.

On-device agentic AI via Snapdragon chips reduces latency and enables offline-capable agent deployments.

Deployment / data residency

All Hands AI launched OpenHands Cloud, a managed SaaS offering of the open-source OpenHands coding agent (formerly OpenDevin), providing a hosted runtime without requiring self-hosted infrastructure.

Cloud availability removes the main adoption barrier for teams that want OpenHands' open-source flexibility without managing the execution environment themselves.

Workflow orchestration

StackAI January release added human-in-the-loop, custom RBAC, SDLC environments, node run results, OpenAI web search, and multiple integrations.

Human-in-the-loop and RBAC in a single release indicate serious enterprise workflow maturity.

Funding / partnership

Decagon announced Series D / $250M commitment.

Significant funding validates category leadership in AI customer-support agents and signals long-term vendor stability.

Product / Positioning

Mastra launched as a TypeScript-first agent framework with built-in support for workflows, RAG, evals, memory, and an agent development Studio UI — positioning as the TypeScript equivalent of LangChain for teams building production agents in Node.js environments.

TypeScript-native agent frameworks are underserved; Mastra's built-in eval and observability tooling makes it relevant for JS/TS engineering teams evaluating agent platforms.

MCP / tool calling / API

Zapier launched an MCP server exposing its 7,000+ app integrations as tool-callable actions for AI agents, allowing external agents (Claude, Cursor, etc.) to trigger Zapier automations directly via the Model Context Protocol.

An MCP server covering 7,000+ apps dramatically expands the integration surface available to any MCP-compatible agent — a major interoperability development for buyers evaluating tool-calling breadth.

Funding / partnership

Lyzr announced $8M Series A and Accenture investment and collaboration.

Accenture involvement signals enterprise-channel distribution and adds credibility for regulated-industry deployments.

Verification status update

AGI Inc. publicly claimed AGI-0 reached 97.4% task success on the AndroidWorld benchmark.

Benchmark performance claims at this level, if verified, would represent a significant capability advance for computer-use agents.

Integrations

Julian operates across voice, chat, SMS, and WhatsApp in IBM Agent Connect materials.

Multi-channel coverage (voice, chat, SMS, WhatsApp) broadens Julian's deployment surface for inbound sales workflows.

Product / Positioning

Phidata rebranded to Agno and released v1.0 of its agent framework, repositioning as a high-performance, model-agnostic Python runtime for building multi-modal agents with native support for memory, tools, and multi-agent coordination.

The rebrand and v1.0 milestone signal production readiness and a sharper positioning around performance and multi-modal agent architectures.

Memory / state

Voice 2.0 added faster voice processing, self-serve controls, and cross-channel memory.

Cross-channel memory enables continuity across voice and text support interactions, improving customer experience quality.

Integrations

Dify 1.7.0 added OAuth support for tools and multi-credential management.

OAuth for tools and multi-credential management enables enterprise-grade secure integration patterns.

Pricing / packaging

Relevance AI split pricing into Actions and Vendor Credits, enabled BYO API keys, removed markup on vendor costs, and sunset the Business plan.

BYO API keys and removal of vendor cost markup can substantially reduce total cost for high-volume workloads.

Agent capability

n8n added native AI Agent nodes with memory, tool-calling, and MCP client/server support to its workflow builder, enabling agentic loops to run inside automation workflows without requiring separate orchestration infrastructure.

Native agent nodes with MCP support position n8n as a hybrid automation-plus-agent platform, relevant for buyers wanting agentic capabilities within existing workflow tooling.

Funding / partnership

AGI Inc. announced a Visa partnership for agentic commerce and secure agent-led transactions.

A Visa partnership signals production-grade agentic payment workflows, relevant for commerce and fintech use cases.

Deployment / data residency

Dify Enterprise became available in Microsoft Azure Marketplace.

Azure Marketplace availability simplifies enterprise procurement and cloud-native deployment.

MCP / tool calling / API

Dify 1.6.0 added built-in two-way MCP support, allowing users to call MCP servers from Dify and expose Dify apps as MCP servers.

Two-way MCP support is a major interoperability signal for builders evaluating agent platforms with tool-calling and MCP requirements.

Browser/computer use

Simular introduced Simular Pro as its first production-ready computer-use agent.

A production-ready computer-use agent tier is a significant maturity milestone, relevant for teams evaluating browser and desktop automation.

Observability / auditability

Dify 1.5.0 added real-time workflow debugging with saved node outputs and live variables.

Real-time debugging with persistent node outputs significantly reduces cycle time when building and maintaining complex workflows.

Observability / auditability

Decagon launched A/B experimentation and Watchtower always-on QA monitoring.

Built-in experimentation and continuous QA reduce the engineering overhead of operating production support agents safely.

Integrations

Added Hunter, ContactOut, Prospeo integrations, and an Outlook Calendar trigger.

Sales data enrichment integrations and Outlook support expand Lindy's coverage for B2B outreach and productivity workflows.

Workflow orchestration

Relevance AI launched Invent, Workforce, and AgentOS — a suite positioning the platform as a full agent operating system.

An AgentOS framing with Workforce management signals enterprise readiness for managing fleets of agents across organizational workflows.

Funding / partnership

11x announced an IBM partnership to distribute Alice and Julian through IBM Agent Connect.

IBM distribution significantly expands 11x's enterprise reach and adds procurement credibility for large-org buyers.

Agent capability

11x launched Julian, an inbound sales digital worker, alongside the existing outbound agent Alice.

Adding an inbound-focused agent completes the full sales coverage story and broadens the platform's addressable use cases.

MCP / tool calling / API

Lindy added a Code action that runs custom Python or JavaScript directly within workflows.

Custom code execution within workflows enables advanced automation without external API calls or separate compute infrastructure.

Pricing / packaging

Devin 2.0 introduced a lower-cost entry point. Current pricing shows Core, Team, and Enterprise packaging.

A clearer tier structure with a lower entry price makes Cognition more accessible to smaller engineering teams.

Agent capability

Devin 2.0 introduced an agent-native IDE experience.

An agent-native IDE changes the developer workflow model, relevant for engineering teams evaluating AI coding agents.

Verification status update

Lyzr's Responsible-AI materials document explainability, auditability, bias control, and compliance posture.

Formal responsible-AI documentation is a prerequisite for enterprise procurement in regulated industries.

Human approval / guardrails

Relevance AI added Slack-based approval for agent tool requests.

Slack-native approval flows make human-in-the-loop controls accessible without switching tools, reducing friction for enterprise adoption.

Product / Positioning

deepset released Haystack 2.0, a complete framework rewrite introducing a component-based pipeline architecture, native support for agentic loops with tool use, and improved composability for RAG, extraction, and multi-agent workflows.

The 2.0 rewrite is the relevant baseline for any current Haystack evaluation — the architecture, APIs, and agent capabilities differ substantially from v1.