The Agentic Index Brief
November 1 to November 15, 2026 · Published November 15, 2026
The week in one line
“Never merges” lasted about two weeks. Hoplite, Sourcegraph and CodeRabbit each shipped a way for the agent to press merge, with conditions attached. Replit Agent learned to check out, Fly.io Sprites and Kernel handed it a card that works exactly once, and Devin started setting its own alarm. The agent gets to finish the job now. The fine print shipped in the same release.
This edition of the log adds 427 entries across 387 vendors, dated November 1 to November 15. Of those, 304 are Verified against a primary source and 123 are Partially Verified. Plain agent capability led with 138 entries, then MCP and tool calling with 74. Human approval and guardrails came third with 50, up from fifth place last time.
That ranking is a fair summary by itself. Agents were handed more to do, and somebody had to write down the terms.
“Never merges” lasted two weeks
In the last issue we wrote that the phrase to look for in release notes was “never merges”. It was a good line. It held up for about two weeks.
Agents in Hoplite threads and swarms can now merge, queue or dequeue their own pull request when the user asks. The agent acts only once checks pass, review threads are resolved and GitHub reports the pull request mergeable. Branch protection still applies, and agents cannot use admin overrides.
Sourcegraph's Agentic Batch Changes can merge the changesets it opened. Nothing merges until the user clicks Allow on a card that lists the problem changesets first, the ones with merge conflicts or failed checks.
CodeRabbit took the person out of the routine case. Teams describe in plain language what is risky in their codebase and what is routine. CodeRabbit then approves the routine pull requests and holds back the ones the instructions flag, and a Triage rule can merge what was approved.
The risk tables moved to match. Factory's Droid now counts a plain push or pull request as medium risk, with force pushes still high. Devin Review can add GitLab merge requests to the merge train. And Google Antigravity's CLI now starts sessions in Auto mode by default, though a strict Tool Permission setting overrides it.
The same thing happened outside the code editor. SAP's new expense agent for Concur tells a manager whether to approve, send back or review closely. When configured, it can approve low risk reports on the approver's behalf. ServiceNow's AIOps AI Specialist can close noisy alerts on its own, with admins setting how much autonomy it gets. When someone calls out of a shift, Rippling's scheduling agent suggests a replacement, and customers can let it offer the shift to eligible coworkers automatically.
Not everyone moved the same way. Claude Code 2.1.325 moved sandboxed Bash commands onto the normal permission flow, so they ask unless a setting says otherwise. Codex CLI 0.169.0 keeps project configuration and MCP servers disabled in untrusted directories until the user passes --trust on purpose. CodeGPT narrowed “Always allow” to the exact command that was approved.
Two more for the collection. Brev's Operator now asks for a plain yes before it changes anything once it has read content written outside the company, so a web page can no longer say yes for you. And in Charm's Crush, YOLO mode now blocks commands classed as dangerous outright. Even YOLO mode has a list of things it will not do.
Our read: we will take the correction. “Never merges” was the right phrase for October. The phrase for November is “when the user asks”. In each of these releases the agent got the last step, and the vendor wrote down what has to be true first. Someone asked. The checks passed. The branch rules still apply. The merge button is the same everywhere. The list of conditions in front of it is where the products differ.
The agent got a card that works once
Last issue the password manager stood between the agent and the login. This time a wallet stood between the agent and the checkout.
Replit Agent can now research and compare products, prepare a merchant checkout through a browser subagent and complete the purchase with Link, Stripe's digital wallet. The user approves the merchant, the total and any recurring terms in the Link flow. Connecting the wallet does not authorize spending by itself, and changed terms trigger a new approval. It is open to Core and Pro users in the US.
Fly.io Sprites added a Link connector that is even more specific. An agent inside a Sprite requests a one time virtual card for a named merchant and amount, and the user approves or denies it in the Link app. An approved card is single use, capped at that amount and invalid after 12 hours. The gateway also blocks the routes that would let the agent approve its own request, a loophole someone clearly thought of before the agent did.
Kernel vaults take single use cards now too. A backend or card issuer sends the card to a Kernel wallet, and Kernel types it into the merchant checkout, so the agent driving the browser never handles the details. OpenRouter let agents pay for their own upkeep. Two new endpoints let an agent top up an OpenRouter account through the Machine Payments Protocol, paying in USDC on Tempo or, where offered, with a card token.
Merchants got a way to check who is behind the agent. Sardine's Know Your Human ties a shopper's own device to an AI shopping agent's session at checkout. The merchant sends the customer a short lived link. The customer opens it and gets a one time code, and the agent submits that code at checkout.
Voice agents took payments as well. ASAPP's GenerativeAgent can collect a card payment during a voice conversation, and Slang.ai's Superhost takes a credit card by phone to book OpenTable Experiences. On the consumer side, Pally now asks before automations, background work or group chats buy something or raise a spending limit.
Our read: for once the expense policy shipped on the same day as the card. Look at the shape of it. One merchant, one amount, one use, twelve hours. That is a permission slip that happens to clear. Stripe's Link now sits where 1Password sat two weeks ago, as the third party that holds the sensitive thing and asks the human. Agents are going to be customers, and the plumbing assumes they will be supervised ones.
Your agent can now call their agent
Personal agents got a front door to knock on. Sierra and Meta published version 1.0 of the Personal Agent Protocol, known as Poppy. It is an open standard for how a personal agent connects to a company, whether through the company's own agent, its website or its APIs.
The mechanics are simple on purpose. A business publishes a poppy.json file that lists its website, APIs or MCP server and its own agent. The customer grants scoped access through OAuth on the business's own sign in page, so the customer decides what the agent may do. A commerce extension coauthored by Shopify makes the agent ask the customer before certain actions.
Adoption came bundled. Every Shopify store now publishes a Poppy file automatically. Muse is the first personal agent to support the protocol, and it already connects with Mercury through Sierra to check balances, categorize spending and request payments or transfers. Turkish Airlines is live as well.
The phone line got the same treatment. AgentID from Scaled Cognition listens to live call audio and detects when a personal AI agent is calling instead of a person. It verifies who the agent represents and what it may do, then routes the call by the business's policy. It sits in front of an existing IVR or voice agent.
Businesses did not wait for the knock. Decagon's Missions let its agent pursue a goal stated in plain language across people, channels and systems, over days, weeks or months, keeping a case record of the steps taken. Lorikeet's Dreamweaver gives each customer a long running AI account manager. It wakes on a schedule or on events like a usage change, and reaches out on the customer's preferred channel.
Our read: robots.txt told crawlers where not to go. A Poppy file tells agents where the front desk is. With every Shopify store publishing one by default, a brand no longer gets to wonder whether personal agents will show up. The open question is what its own agent says when they do. Somewhere soon, two agents will settle a refund while both humans are at lunch.
The agent set its own alarm
An agent that waits for a prompt is a tool. Since November 1, a lot of them stopped waiting.
Devin can now set itself reminders to wake up at a future time, and the session shows a Reminder set row with the time and a link. Hoplite agents can schedule a delayed follow up for their own thread, set recurring wakes and watch a command until it succeeds, each with a visible status the user can cancel. camelAI's camelRun SDK added monitors, a check script that runs on an interval and wakes the agent only when there is something to do.
For everyone who does not live in a terminal, the same idea arrived in plain language. Lovable added automations, where a user asks for a recurring task in chat, Slack or Telegram and says how often. Unblocked launched Routines, instructions that run on a schedule, when a Linear issue is created or when a message lands in a chosen Slack channel. Zendesk's custom agents, rolling out through November 19, can be invoked on an event or a recurring schedule.
Events started reaching agents directly. MCP events in the Chatbase MCP server let a connected AI client run a standing task whenever a new ticket arrives, a conversation is escalated or a phone call ends. Letta Code agents can manage durable event subscriptions that stay active after the agent's computer stops. Wassist uses the same MCP Events extension, so an order shipping can prompt the agent to message the customer on WhatsApp.
Work that runs unattended needs a meter, and those shipped too. Letta Code's Workflow tool takes a maxCostUsd budget and stops the run when it is reached. Every Enclave workflow run carries a credit budget it will not exceed. Middleware lets admins cap automatic investigations by day, week or month. And ContraForce put hourly limits on response actions, so a misconfigured run cannot isolate every endpoint at once.
Our read: scheduling sounds like a small feature. It is also the line between an assistant and staff. Once an agent can wake itself, the useful questions change to what it plans to do at 3am, who asked for that, and where the off switch is. The better releases answered in the interface, with a visible status, a cancel button and a budget. The alarm clock is fine, as long as someone can see what time it is set for.
The swarm got a manager
Running one agent is a prompt. Running twelve is management, and the tools started to admit it.
Cursor Projects, where one coordinator directs a fleet of worker agents, is now multiplayer. Teammates join a Project with their own agents and share a task board. Their coordinators talk to each other to spot overlapping work and agree on order.
Google Antigravity added /swarm in public preview. The agent becomes a coordinator, each worker gets its own conversation and Git worktree, and a monitor worker alerts the coordinator when the swarm stalls. Artisan's AI employees can now see who else has picked up a request and leave the overlap alone. Junie CLI from JetBrains lets a developer message a subagent while it works instead of waiting for its report.
Atomicwork went furthest with the metaphor. Six certified AI Coworkers are in its marketplace, each with tested connectors, an operating procedure, a reporting manager and a budget. That is a more complete onboarding packet than most people get.
Our read: the org chart is being rebuilt one layer down. There are coordinators, workers and a monitor that notices when everyone has gone quiet. The names are borrowed from management because the problems are. Two agents doing the same task is a Monday standup problem, and no model upgrade fixes it.
Market notes
The kill switch became standard equipment. CrowdStrike's Falcon Guardian added an AI Kill Switch that lets an administrator terminate a running agent, block it from relaunching and cut its network access, working at the AI Gateway today. Stacklok put one on every gateway. Tuskira's open source gateway can move an agent's key to restrict or to kill. LangSmith lets org admins lock an agent, which stops new runs. Add Salesforce from the last issue and the kill switch is starting to look like a checklist item.
The agent moved onto the desk. n8n Labs released Nod as a public research preview, an open source personal agent for Mac and Windows with its own local sandbox and browser. Optimizely's Mark Desktop gives its agents local file tools and shell commands, limited to folders the user grants. Perplexity Computer can run entirely on device on supported Apple silicon Macs. And Microsoft's Copilot Cowork can use the local Edge browser in a separate tab, with credentials and cookies kept on the device.
Agents started joining meetings. Chipp's assistant Victoria joins Google Meet calls she is invited to and saves a summary with decisions and action items. Grok Bot joins a Google Meet, Zoom or Microsoft Teams call from a link and answers in the meeting chat. Salesloft went the other way and now records Zoom meetings through Zoom's media streams, so no bot appears in the participant list. The meeting of the future has either one more attendee or one fewer.
Two releases bet that agents work better writing code than clicking. Claude Code 2.1.323 added Code Mode, an experimental opt in where Claude writes JavaScript that calls its tools instead of requesting each call one at a time. Browser Use 1.0 replaced its engine along the same lines. The agent now writes Python against the browser rather than picking one action per step.
App builders pointed their agents at their own output. v0 can run a security review on request and only changes code if the user agrees. Zite added a Security page where an agent lists findings by severity and offers Fix or Fix all. Lovable now asks users to review critical security findings before publishing. Base44's scan looks for prompt injection and for ways someone could run up the bill on the builder's own AI key.
Voice kept getting quicker. Retell released Tell-1, a model it trained in house that starts speaking right away and reasons while its first sentence plays, at $0.035 per minute. Hippocratic AI previewed Polaris 6 Ultra and reports 278 ms to first sentence. Twilio made Smart Turn Detection generally available, replacing silence timers with a model that decides whether the caller is finished or only pausing. And Pipecat's Voice Reader reports the caller's tone, laughter and sarcasm, so the agent may finally know what you meant by “great, thanks”.
In the back office, Workday's Travel Agent went generally available. It books flights, hotels and rental cars in one conversation and checks each option against company policy. FurtherAI's agents now run insurance submissions through quote and bind, with underwriters approving the risk and setting the final price. Factorial's agent can read a photo of a handwritten shift schedule and apply the differences once the user confirms.
On price, Salesforce added a $50 per user per month SKU for Agentforce Coworker. Atlassian renamed Rovo Dev code review to Code Reviewer and included it with any paid Jira or Teamwork Collection license, with credit billing starting December 3. Amp launched boxes, always on VMs for the apps its agent builds, from $18 per 30 days. And R1 completed its acquisition of Humata Health and relaunched its prior authorization technology as Phare Auth.
One small one to finish. Letta Code made apps exposed from a cloud sandbox private by default, and its skill instructions now tell agents never to add the public flag just to avoid a sign in. Anyone who has managed an intern will recognize the memo.
What the period says about the category
Last issue the question was who lets the agent in. This time it was what the agent is allowed to finish.
It can merge the pull request, once someone asks and the checks pass. It can pay, with a card good for one merchant and one amount. It can introduce itself to a business that has published where to knock. It can wake itself up, on a meter, with a cancel button.
Each of those is a last step that used to belong to a person. None of them shipped bare. The conditions arrived in the same release as the capability, which is not how software usually goes.
The agent is still the new hire. It now has a company card with a low limit, a calendar and permission to close its own tickets. Nobody has given it the admin override, and Hoplite put that in writing.
Which AI SRE agents can resolve an incident on their own?
About four in ten document the whole loop. The Agentic Index grades 48 vendors in its SRE and DevOps lane. Resolving an incident without a person driving takes five capabilities, each documented in full. A signal has to start the agent. The agent has to reach the stack for evidence. It has to run multi step work rather than one lookup. It has to leave an audit record. And a human has to govern what it may change in production. 20 of the 48 clear all five.
Getting started is the solved part. 43 of 48 document triggers in full and 41 document integrations. The count drops to 36 for multi step orchestration, 32 for the audit record and 30 for human oversight. So most agents in this lane can take the page off the queue. Fewer can show their work afterwards, and fewer still document who gets to stop them.
Another 11 sit one capability short. 4 are missing only the oversight evidence, 3 the orchestration, 3 the audit record and 1 the trigger. Across the lane, 6 vendors document the brakes without the flight recorder, and 8 document the recorder without the brakes. Memory is rarer than either, at 10 of 48, which matters in a job where the same outage tends to come back.
The releases since November 1 went straight at those gaps. PagerDuty took its SRE agent Paige to general availability. Teams can keep a human approving remediation or let Paige resolve well understood incidents within boundaries they set, and it writes each resolved incident back into shared memory. The Dynatrace SRE Agent, also generally available, prepares a pull request, config change or runbook step that waits for human approval. ServiceNow's AIOps AI Specialist records its decision rationale and evidence, and admins set its autonomy level. groundcover's Investigations keep the query, source and time range behind every finding. Oversight and the audit trail are turning up as product features, where a buyer can see them in a demo.
The bar, the method and the vendors that clear it are on the Agentic Index production ops agents page.
The Agentic Index Brief is published weekly by Agentic Index, the verified directory of 1,167 agentic AI vendors. Compare platforms by capability at agenticindex.io/compare. Methodology at agenticindex.io/methodology.