Release gate checklist for AI agent teams: what counts as proof
A merge and release gate checklist for teams shipping with AI coding agents: required checks, skipped jobs that report green, the security verdict on the exact commit, and the human promotion step.
How to use it
- Merge gate on every pull request. Release gate before every promotion to production.
- A box is ticked with evidence (a run id, a URL, a query output), never with a sentence.
- The person who promotes is never the agent that built it, by construction: no agent holds the credentials the workflow uses.
# Merge gate (every pull request)
- [ ] Lint passes, including the house rules.
- [ ] Types pass in strict mode.
- [ ] Unit, integration and database tests pass. Run id: <id>
- [ ] Secrets scan passes on every commit of the branch, not only the last.
- [ ] Browser tests RAN on this change. Opened the run: <n> steps executed,
<n> tests passed. A skipped job reports as success; it is not a pass.
- [ ] The browser job filter lists what may skip it, not what triggers it.
- [ ] On any change to database security: one isolation rule was broken on
purpose and the suite went red. If it stayed green, the tests do not
test what they claim.
- [ ] Red checks read from the annotations first. Seven failures in four
seconds are usually one cause, not seven bugs.
# Release gate (every promotion to production)
- [ ] All required checks green on the exact commit being promoted, not an
earlier one. Commit: <sha>
- [ ] Security agent penetration test against local and staging (never
production) says PASS for that commit. Report: <link>
A FAIL blocks the promotion like a red check, even if it is outside
the feature being shipped.
- [ ] Data compatibility queries run before promotion. Output: <link>
- [ ] No known failure without a ticket. A comment saying "escalated" is
not an escalation.
- [ ] Promotion is one human command, logged with who and when.
- [ ] The approver is not the agent that built it.Prompt to use it
You are reviewing pull request <number> against the merge gate in docs/release-gate.md.
For every box, report PASS or FAIL with the evidence: a run id, a step count, a URL or a query output.
A job that was skipped is FAIL, not PASS. Say which jobs were skipped and why.
If the change touches database security, break one isolation rule on purpose in a scratch branch,
run the suite, and report whether it went red. If it stayed green, that is the most important finding of the review.
Do not write "looks good". End with: VERDICT: PASS or VERDICT: FAIL, and the boxes that failed.Why it looks like this, and where each part came from: Why a green check is not proof.
Changelog
- v1.0Nov 5, 2026First public version. The skipped check rule dates from a navigation redesign merged with eight green checks while the browser tests had not run; twelve went red on the next pull request.
More from the kit
CLAUDE.md template
The seven sections every agent session inherits, in the order an agent reads them, with nine neutral hard rules to start from.
v1.0 / CLAUDE.mdFormat / BuildEscalation and decision format
How an agent stops at a real decision: an id, three options, a recommendation. You answer with a letter.
v1.0 / escalation.mdTemplate / BuildTicket template for agents
Frontmatter a script can validate, plus the three fields that stop agents inventing requirements. Blank and filled versions.
v1.0 / ticket.mdWorking on something like this?
This is the method behind how I build: AI native development with human gates. Discuss a product idea → Or email me.