Your agent just said "All tests pass." Ship it?

Not so fast. When one agent writes the code, the tests and the verdict, what is left to rely on? Green is the agent's finish line, not a go-ahead, and certainly not proof.

Coverage can't tell you much, since agents hit any coverage target without checking anything. Mutation score can, because it asks whether your tests catch faults injected on purpose. The tests that catch them are specification tests: written from the requirements, not from the code, they say what the software must do, whoever writes the implementation.

Then you'll see how to have tests written by an agent that has never seen the code, and how to keep them Readable, Enforced and Durable through every rewrite.

Live, in a Java project, you'll watch a coding agent try to weaken its own tests to reach green, and watch the guards stop it.

- Write specification tests from the requirements, separate from the scaffolding tests agents write for themselves.
- Measure test quality with mutation score**, not coverage, and spot tests that can never fail.
- Stop agents from weakening, skipping or deleting the tests they're judged by.
- Set up the skills, hooks and CI gates that keep tests trustworthy.