Clearing a code review backlog

Small teamsTeam of five with 20+ open pull requests

We had 23 PRs waiting. Reviewers asked their agents 'what does this change and why', and each agent re-read the diff and asked the same questions the author's agent had already answered while building it.

How can reviewers' AI agents see why a pull request was built the way it was?

Have the author's agent save decisions and failed approaches with decide and remember_failure while it builds, and ArcRoom hands them to the reviewer's agent when it asks about the PR. A question like 'why not use a queue here?' comes back with the failed attempt and the reason. It only helps when the author's agent recorded its reasons. ArcRoom doesn't read PR comments.

Also asked as: Our PR queue is huge and reviewers' agents keep re-asking the same questions. What helps? · How do I share the reasons behind a code change with the reviewer's AI agent? · Can a reviewer's AI agent see approaches the author's agent already tried and dropped?

The work

While building, the author's agent saves decisions and failed approaches with decide and remember_failure.

ArcRoom keeps
  • Failed approach
  • Decision
  • Saved decision
  • Recall
The next agent gets

Recall for PR #418 (author: neha):

Step by step

  1. While building, the author's agent saves decisions and failed approaches with decide and remember_failure.
  2. The reviewer's agent asks about the PR and recall hands over why it was built that way.
  3. 'Why not use a queue here?' comes back with the failed attempt and the reason.
  4. Review time goes to real problems.

What the agent sees

Recall for PR #418 (author: neha):
Tried and failed: BullMQ for exports. Why: Redis not available on the worker host.
Decision: exports run as a cron job, max 10k rows.

Compared with today's tools

A PR template with 'why' and 'alternatives tried'

Authors fill in a template on every PR, and reviewers paste it into their agent.

Setup
30 min
Ongoing
authors write it up per PR, ~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

A good PR description is better for human reviewers; ArcRoom's edge is that the reasons are captured even when the author skipped the template.

Where it stops

Only helps when the author's agent recorded its reasons. ArcRoom doesn't read PR comments.

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