A verification checklist for agents

1. What you define before the task

Autonomy does not start with permission. It starts with a definition of done you can check.

• What the task is supposed to change.
• What it must not change.
• What evidence proves the task is finished.
• Who or what checks the evidence.
• What happens when the check fails.

2. Four basic kinds of evidence

Not every task needs the same check. Pick the evidence based on the risk.

• Code: a targeted test, typecheck, lint, or a build that passes.
• Interface: a before and after screenshot, on mobile and desktop.
• Data: one query that returns an expected number or sample record.
• Content: a duplication check, a CTA that does not clash, and a live link if an asset is promised.

3. A review agent

If the agent that built it also approves it, you get the same bias twice. Give a second agent read-only access.

• Read-only permissions.
• Goal: find a problem, not fix it.
• Output: problems with a file, line, screenshot, or check attached.
• No code changes and no external action.

4. A prompt to paste

Before you say done, show proof.

Return:
1. Files changed
2. Tests or checks run
3. Evidence link or screenshot
4. Risks left
5. What you did not verify

If proof is missing, say not done.

5. What to do when the check fails

• Do not rerun the same request unchanged.
• Isolate the problem: code, interface, data, or content.
• Narrow the permissions until there is a small piece of evidence that works.
• After two failed attempts, open a review or hand it back to a person.