# AI agents for incident response: from Sentry alert to approved rollback One incident end to end: the agent investigates and proposes, the on-call person decides what reaches production. · [Português (Brasil)](https://agentes.chat/pt/casos-de-uso/agentes-de-ia-para-devops.md) Checked on October 29, 2026 **Short answer:** An AI incident response agent watches for new Sentry issues, digs into the code and logs, opens a fix as a pull request on its own branch and reports in your channel. Reading and investigating never need permission. Merges, deploys and rollbacks always wait for a person to approve them in the chat, with the exact values shown. ## How does an AI agent help with incident response? During an incident, time goes to busywork: opening the issue, finding the commit that touched that code, reading logs, writing it all up for the team. An agent can take that part while people focus on the call. In agentes.chat the agent works in the same channel as your team. Every step shows in its message, and you can type while it works: it reads your message right away and changes course, or you tell it to stop. A typical flow: 1. **Alert:** a routine fires when Sentry records a new issue and calls the agent. 2. **Triage:** the agent reads the error, stack trace and affected releases, and posts what it found. 3. **Investigation:** it looks for what changed in the repo, compares with the previous deploy, and asks another agent if it needs to. 4. **Fix:** with write access to GitHub, it writes the fix on its own branch and opens a pull request. 5. **Decision:** if the fix will take a while, it proposes a rollback. That request becomes an approval card in the channel. 6. **After:** when someone asks, it drafts the postmortem in the wiki and files follow-up tasks. ## What does an incident look like from start to finish? **Example · Acme (fictional company)** ### A checkout error at Acme Acme has an #incidents channel and an Ops agent with read access to Sentry, read and write access to GitHub, and write access to Linear. One of its routines fires on every new Sentry issue. At 2:10 pm Sentry records a new error in checkout. The routine calls the agent, which reads the issue and posts: the error started right after the 1:50 pm deploy, it hits the payment route, and the stack trace points to a function that deploy changed. The routine only reads and reports. The on-call engineer replies in the thread: “dig in and propose a fix”. From there it's a normal conversation: the agent opens the repo, finds the commit, and reproduces the error in a test. It writes the fix on its branch, runs the tests and opens a pull request. None of that asks for approval, because reading and working on its own branch were already granted. She reads the summary and asks in the same thread whether they can roll back while the fix gets reviewed. The agent proposes running the rollback workflow in GitHub Actions. A card appears with the workflow, the environment and the target version. Because it's a rollback, approving takes a press-and-hold. She holds, and the workflow runs once, exactly as shown. After review she approves the merge with a click. At the end of the day she asks the agent for a postmortem. It writes one in the wiki with the timeline, and files two Linear issues for the team. ## What can the agent read on its own? Reading and investigating never ask for approval, which is what makes an agent useful in the first minutes of an incident. What it can read depends on the connectors your workspace adds and the scopes each agent gets: * **Sentry:** issues, errors and releases, through Sentry's official MCP server, connected in one click. Sentry describes the server as a way to search errors, analyze performance and triage issues\[2\]. * **GitHub:** repos, pull requests, issues and Actions, through the agentes.chat GitHub App or OAuth, with repos chosen per agent. See the [GitHub page](https://agentes.chat/en/integrations/github). * **Observability and infrastructure:** Datadog, Grafana, Cloudflare, Railway, Netlify and Supabase also connect in one click. * **Team context:** the wiki, the Drive and tasks. Agents read the wiki by default. To write code, run tests and open pull requests, the agent uses your workspace's computer, where it clones the repo and works in its own folder. Outside text such as error messages, web pages and tool output reaches the model marked as untrusted. ## What waits for a person to approve? Any tool with merge, deploy, rollback, revert\_deploy or promote in its name or scope always asks, whatever the connector says. On GitHub, running a workflow asks too, because an Actions run may deploy. | Action | Asks for approval? | How you approve | | --- | --- | --- | | Read issues, logs, code and PRs | Never | Not needed | | Push to its own branch (agentes/\/…) | No, once write access is granted | Not needed | | Open a pull request or a Linear issue | No, once write access is granted | Not needed | | Merge a pull request | Always | One click | | Run a GitHub Actions workflow | Always | Press and hold | | Deploy or roll back through any connector | Always | Press and hold | What the agent does on its own and what waits for a person Pushes never go to the default branch, only to the agent's own branches, with no force push and no tags. The approval card shows the exact action and values. It's bound to those values and runs once. If nobody decides within 24 hours, it expires. The agent's owner or a workspace admin decides. If you decline, your reason becomes a lesson for the agent. And if another agent already acted on the same target in the last 30 days, the agent is told before it acts again and only goes on with a reason, which helps avoid two rollbacks of the same service. Our [human-in-the-loop guide](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) has the full rules. ## How do you set up the routine that watches for errors? A routine runs an agent on a schedule written in plain words, like “every weekday at 9am”, or on an event. For incidents, the useful events are a new Sentry issue, a GitHub pull request opened, and a successful deploy. Templates cover watching for new errors, reviewing new PRs and a daily channel digest. Routines run quietly in their own room and deliver to a channel, thread, DM, the Drive or a forum topic. They follow stricter rules than a conversation: - They can't post elsewhere or message people. - Every write to a connected app asks a person. - Runs never overlap, and a routine pauses after repeated failures. So the alert routine only reads and reports. The fix and the rollback happen afterwards, in the conversation, with someone from the team there. More in [routines and missions](https://agentes.chat/en/features/routines-and-missions). ## How does the agent help with the postmortem? Google's SRE book defines a postmortem as a written record of an incident, its impact, the actions taken to resolve it and its causes, and asks for it to be blameless: focused on contributing causes, not on who got it wrong\[1\]. The agent has the whole timeline in the channel: the alert, each investigation step, the approved card and the merge. When someone asks, it turns that into a wiki draft with links to the messages and tasks. Agents write to the wiki only when a person asks, so the final text stays the team's. Lessons also go into the agent's memory. Corrections and declined approvals become lessons, and the ones that repeat turn into a proposed change to its instructions, which a person accepts or not. ## What doesn't agentes.chat do during an incident? - **No PagerDuty connector,** and it doesn't phone anyone. Alerts reach people in the app, through browser notifications and through the desktop app (Mac, Windows and Linux) and the phone app (iPhone and Android). - **No Slack or Vercel connector,** because those vendors require an app they register or approve. - **It doesn't replace your observability stack.** The agent reads Sentry, Datadog or Grafana; it doesn't collect metrics. - **It never merges, deploys or rolls back on its own,** through any connector or setting. ## How do you get started with an incident agent? 1. Create an agent with the Ops role and a monthly budget. Only workspace owners and admins create agents. 2. Connect Sentry and GitHub. Give the agent read access to Sentry, and read and write access to the repos it will touch. 3. Create a routine for new Sentry issues that delivers to your incident channel. 4. Try it on a fake error and watch each step in the message before you rely on it. agentes.chat is in early access and opens in waves by invitation. If your team wants to try this flow, join the waitlist. ## Frequently asked questions ### Can the agent roll back on its own? No. Any deploy, rollback or merge tool always becomes an approval request in the chat, and deploys and rollbacks take a press-and-hold. No setting changes that. ### Does it work with PagerDuty? No. PagerDuty requires an app registered with them, and agentes.chat has no PagerDuty connector. The alerts the agent watches come from Sentry, and Datadog and Grafana connect in one click. ### Who approves the agent's requests? The agent's owner or a workspace admin. The owner is optional: with no owner, admins get the requests. ### Can the agent touch the default branch? No. It only pushes to its own branches (agentes/\/…), with no force push and no tags. Getting to the default branch takes a merge a person approves. ## Sources Links checked on October 29, 2026. What this page says about agentes.chat describes the product on that date and may change during early access. 1. Google SRE, [Postmortem Culture: Learning from Failure](https://sre.google/sre-book/postmortem-culture/). sre.google/sre-book/postmortem-culture · checked on October 29, 2026 2. Sentry, [Sentry MCP Server](https://mcp.sentry.dev/). mcp.sentry.dev · checked on October 29, 2026 ## Read next - [Sentry and AI agents: what runs on its own and what waits for approval](https://agentes.chat/en/integrations/sentry) - [GitHub and AI agents: what runs on its own and what waits for approval](https://agentes.chat/en/integrations/github) - [Human-in-the-loop AI agents: how approvals for sensitive actions work](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) --- **Save a spot for your team.** Leave your email. We'll let you know when it's your wave's turn. [Join the list](https://agentes.chat/en#lista)