A green test suite can still approve the wrong behavior. Acceptance testing checks a feature against product requirements, independently of what its implementation appears to do. That independence matters when an agent writes both the code and its tests: both can encode the same misunderstanding. section 10.4.1 section 10.4.2
A rounding error that passes
A test expected a 50% discount on 9.99 to return 4.99, and the code returned 4.99. Both agreed, but the required answer was 5.00. An acceptance test derived from the requirement rather than the implementation can expose that mismatch. section 10.4.1
Agent-generated unit tests can help detect regressions, but they do not independently establish that intended behavior meets product requirements. Use human-written acceptance tests or tests generated from ticket requirements, not from the code, to provide that separate check. section 10.4.1 section 10.4.2
For agent-generated pull requests that change business logic, require at least one human-written or ticket-derived acceptance test. In one team, this added about 15 minutes per pull request and caught three specification violations in the first month. Check test integrity as well: a green CI result can also come from disabling tests or weakening assertions. section 10.4.2 section 10.4.1




