How do I verify a coding agent's work?

Keep the checks under your control

1. 1. Define

   Write the expected behavior and failure cases before the agent starts.

2. 2. Compare

   Run unchanged acceptance checks against the original code and the candidate.

3. 3. Review

   Inspect the code diff, check output, and any unresolved limits.

For a bug fix, the relevant regression check should fail on the original code and pass on the fix. Checks for an added feature can start with that feature absent. Run unrelated checks too, so a narrow fix does not hide another failure.

What if the agent changes the tests?

Keep a trusted copy of the acceptance suite in a separate verification checkout that the agent cannot edit. Use a disposable environment for candidate code. A separate folder alone does not enforce permissions; check the agent's actual write access.

The agent can propose new tests or explain a faulty test. Review that change against the requirement separately. If the requirement did not change, a weaker assertion is not proof of a fix. I keep both the old and proposed test output with the diff.

Run the example yourself

This synthetic example counts two TODO notes and one DONE note. Broken code counts only one TODO. A weakened test expects that wrong answer and passes. The fixed code passes the original acceptance checks.

I use Python's standard library here. The example makes no model or network calls and changes no files. It demonstrates the test boundary, not a BMD run or a customer outcome.

python verify-agent-work.py

What evidence should the agent provide before it says done?

• The requirement, expected output, and unchanged acceptance checks.
• The exact candidate revision and the code and test diffs.
• The check command, environment, exit status, and complete output.
• Failures, skipped checks, and behavior that remains unverified.
• A browser check for visible changes, or a readback from the actual system when the task changes live state.

Passing these checks proves the behavior they exercise. It does not prove every behavior of the program. A deployment, email, or payment needs evidence from that system; local tests alone do not prove the external result.

Sources and a next step

The question came from a public discussion about agents changing tests. Claude Code's verification guidance explains why concrete checks matter. The runnable example and its output are the proof for this guide.

BMD offers a Windows workspace for agent queues, runs, and review. Check its release notes for the behavior in your version. You still need acceptance checks you trust.