Release captain, a new one every week

Small teamsTeam of eight with a rotating release captain

Each week a different person cuts the release. Last week Kiran's agent hit a migration that had to run before the deploy, and lost an hour. This week Tara's agent hit the same thing.

How do I pass last week's release problems on to this week's release captain and their AI agent?

With ArcRoom, problems during a release go in with remember_failure: what was tried and why it broke. Lasting fixes to the process go in with arcroom decide. When the new captain's agent starts 'cut release', recall hands over past weeks' problems and decisions, and done-checks in .arcroom.json keep it from stopping while the build is red. Only problems someone recorded are passed on.

Also asked as: Every release captain hits the same deploy problem. How do we stop repeating it? · How do I give an AI agent the lessons from past releases? · Can an AI agent remember what went wrong in last week's release?

The work

Problems during a release go in with remember_failure: what was tried and why it broke.

ArcRoom keeps
  • Done-checks
  • Decision
  • Saved decision
  • Recall
The next agent gets

Tried before and failed (Oct 24, kiran): deploy then migrate.

Step by step

  1. Problems during a release go in with remember_failure: what was tried and why it broke.
  2. Lasting fixes to the process go in with arcroom decide.
  3. When the new captain's agent starts 'cut release', recall hands over past weeks' problems and decisions.
  4. Done-checks in .arcroom.json keep the agent from stopping while the build is red.

What the agent sees

Tried before and failed (Oct 24, kiran): deploy then migrate.
Why: new code reads the column before it exists.
Decision: migrations run first, in a separate step.

Compared with today's tools

A RELEASE.md runbook

Keep the steps in a runbook and add a line after each bad release.

Setup
1–2 hours
Ongoing
captain updates the runbook after each release, ~10 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 per person (npx arcroom init)
Ongoing
None

The runbook is the better record of the steps; ArcRoom catches the problems nobody wrote into it, because the agent recorded them as they happened.

Where it stops

Only problems someone recorded are passed on. The runbook still belongs in the repo.

Try it in your project

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

More in small teams