An intern's agent that learns from review

Small teamsEngineering team with a summer intern and a mentor

Our intern's agent kept making the same three mistakes our mentor flagged in review: raw SQL in handlers, no tests for error paths, and console.log left in. Each PR got the same comments.

How do I stop our intern's AI agent from repeating the same code review mistakes?

After each review, the mentor records each repeated point with arcroom decide and the reason, and ArcRoom hands those rules to the intern's agent on matching prompts. Done-checks in .arcroom.json, like lint and tests, must pass before the agent can stop. The mentor sees the intern's sessions and tokens on the team dashboard, though not a live view of the agent.

Also asked as: Why does my intern's AI agent keep making the mistakes I flagged in the last PR? · How do I turn code review comments into rules an AI agent follows? · Can an AI agent be made to pass lint and tests before it says it's done?

The work

After a review, the mentor records each repeated point with arcroom decide and the why.

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

Decision (Jul 2, mentor): no raw SQL in handlers; use the repo layer.

Step by step

  1. After a review, the mentor records each repeated point with arcroom decide and the why.
  2. Done-checks in .arcroom.json (lint, tests) must pass before the intern's agent can stop.
  3. Recall hands the review rules to the intern's agent on matching prompts.
  4. The mentor sees the intern's sessions and tokens on the team dashboard.

What the agent sees

Decision (Jul 2, mentor): no raw SQL in handlers; use the repo layer.
Done-check failed: pnpm lint
no-console: src/api/export.ts(18)

Compared with today's tools

Lint rules + a CONTRIBUTING.md

Turn repeat comments into lint rules where possible and the rest into a contributing guide.

Setup
2–3 hours
Ongoing
mentor updates the guide as patterns repeat, ~15 min a week

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

Lint rules are stronger where they fit; ArcRoom carries the comments a linter can't check, like where SQL belongs.

Where it stops

The dashboard shows sessions and tokens, not a live view of the intern's agent. That's on the roadmap.

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