An agent woken by an alert

CI and automationTeam with an on-call agent triggered by errors

An alert fired: payments API 500s. The on-call agent started from nothing, though the same alert had been fixed twice before.

How can an on-call AI agent check how we fixed the same alert the last time it fired?

After each incident, have the agent save the cause and fix with ArcRoom's complete_task. When the alert fires again, find_task returns the earlier incidents, so the agent checks known causes first and pages a human with a short summary. The agent sees only what was saved; logs and dashboards still have to be read live.

Also asked as: Our on-call agent starts from scratch on alerts we've fixed before. How do we change that? · How do I give an incident response agent a history of past fixes? · Can an AI agent remember the cause of a previous outage?

The work

After each incident, the agent saves cause and fix with complete_task.

ArcRoom keeps
  • Reused answer
  • Saved answer
The next agent gets

find_task → 2 similar incidents: "payments 500"

Step by step

  1. After each incident, the agent saves cause and fix with complete_task.
  2. When the alert fires again, find_task returns the earlier incidents.
  3. The agent checks the known causes first and pages a human with a short summary.

What the agent sees

find_task → 2 similar incidents: "payments 500"
• Sep 12: Stripe key rotated, env not updated
• Oct 3: DB pool exhausted after deploy

Compared with today's tools

Postmortems and runbooks in a wiki or the incident tool

After each incident someone writes a postmortem, and the alert links to a runbook the on-call agent is told to read.

Setup
1–2 hours per runbook
Ongoing
Write a postmortem after each incident, ~30 min

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

Runbooks are reviewed and serve people too, so keep them. An agent reads them only if its prompt points there, while find_task returns past incidents on its own.

Where it stops

The agent sees what was saved. Logs and dashboards still have to be read live.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in ci and automation