See it run

What you just watched

1. You label the issue

   An ordinary GitHub issue gets the agent label. That is the whole trigger. Nothing else changes about how your team files work.

2. A worker on your machine claims it

   The run is claimed before any work starts, and the claim is posted back to the issue, so a second machine picks up something else instead of racing it. The issue moves to agent:running.

3. Classify, branch, plan

   The run decides what kind of change this is, cuts a branch from your base, and writes a spec before touching code. The plan is committed, so you can read what it intended to do and compare.

4. Implement, then argue with itself

   It writes the change, then reviews its own diff against the issue's acceptance criteria and fixes what it finds. Across 73 timed runs this review loop was 45% of total runtime, more than double the implement step.

5. Evidence, attached

   Tests and screenshots are captured and committed with the work, so the PR arrives with proof rather than a claim.

6. It opens the PR. It never merges.

   The run ends at a pull request marked agent:done, awaiting your review. The agent never pushes to your base branch. That part is yours.