Build Your First Loop
Start with one small loop you can personally review. Pick one discovery source, one outcome, one validation step, one checker, one stop rule, and one place to save the run note.
What is the smallest first loop that still produces reviewable evidence without granting unnecessary authority?
What we tested
We constrained one CI repair to a single known failure, one isolated worktree, three attempts, named commands, a read-only checker, and a final human merge decision. We then checked whether the generated artifact preserved every boundary.
Observed result
The narrow setup produced a complete goal without needing repository-wide autonomy. The artifact could prepare evidence and a reviewable branch, yet it had no reason or permission to merge, deploy, access production credentials, or expand into unrelated cleanup.
What this does not prove
The exercise does not determine the ideal retry count for every project and does not measure live agent success. A first loop should inherit the target repository's real tests, permissions, and review rules rather than copying this example unchanged.
Why it matters
The first loop should teach you trust, not replace your judgment. A small CI, review, or content-maintenance loop is easier to inspect, cheaper to run, and safer to improve than a broad autonomous workflow.
A good first loop path
- 01Need a concrete signal?Use CI failures or review comments
- 02Need a safe workspace?Use a branch, worktree, or sandbox
- 03Need proof it worked?Run the same checks a human would run
- 04Need to avoid drift?Write a run note before stopping
Practical checklist
- Choose one discovery source: CI failures, pull request comments, issues, or a small inbox
- Write the goal as an outcome, not a list of instructions
- Add the validation command, score, artifact, or checklist that proves the work is done
- Name the independent checker and what it must reject
- Set max iterations, budget limit, fallback, and human approval before merge or deploy
- Save a short run note: what was seen, what changed, checks run, and what to do next
Example
First loop: every weekday morning, read failed CI on open pull requests. If a failure is small and reproducible, draft the smallest fix in an isolated worktree, run lint/test/build, ask a separate checker to review the diff, then stop for human approval.
Common failure modes
Related templates
FAQ
Not immediately. Run it manually until the checker and stop rule feel boringly reliable, then consider a schedule.