The studio
How should AI agents write pull requests?
Small, one change each, with a plain-English title and a body that says why the change was made, how it was verified and how to roll it back. A reviewer, a person or another agent, should be able to judge it without asking the author anything. That's the shape every pull request at North Bay Ventures follows, whoever wrote it.
Updated November 17, 2026
Why does the format matter more with agents?
Agents open far more pull requests than people do, and often nobody watched them work. The pull request is the record: what changed, what the agent believed was true, and what it checked. A vague one can't be reviewed well, and a big one can't be reviewed at all. See how our AI agent teams ship code.
What goes in the title?
What changed, in plain English, so someone scanning a list of merged changes understands it without opening it. "Sandboxer FAQ: idle auto-pause now defaults to 30–120 minutes" says more than "Update faq.json". The studio's daily reports list merged pull requests by title, so the title has to stand on its own.
What goes in the body?
- Why: the reason for the change, with the source. For a fact, the page or file it came from; for a request, who asked and their words if it needs approval.
- What changed: a short list, especially anything a reviewer might not expect, and anything left alone on purpose.
- How it was verified: the build or tests that passed, the checks run and what was compared. "Looks right" isn't verification.
- How to roll it back: usually "revert this commit", plus any manual step a revert won't undo.
Link the issue if there is one.
How small should it be?
One change per pull request. Small changes are easier to review, merge quickly without conflicts and are easy to revert on their own. When agents work in parallel, small pull requests on short-lived branches are what keep them from colliding; see feature flags vs feature branches and running many coding agents in parallel.
Who reviews an agent's pull request?
At North Bay Ventures, a second agent reviews any change to product code, billing, sign-in or infrastructure before it merges, given the diff and the goal rather than the author's own summary, so it checks the change and not the story. The author fixes what it finds. Content-only changes skip that review, and every change must pass the automated checks. Then the change merges itself and goes to production, where "done" means deployed and checked, not merged.
Frequently asked questions
Should an AI agent merge its own pull requests?
At North Bay Ventures, yes, once the automated checks pass, and after a second agent's review for product code, billing, sign-in or infrastructure.
What should a pull request description include?
Why the change was made, what changed, how it was verified and how to roll it back.
How big should an agent's pull request be?
One change. Small pull requests are easier to review, merge quickly and revert cleanly.
Does a merged pull request mean the change is live?
Not until it's deployed and checked in production. Merged and live are different steps.