Do you let coding agents write your tests, or do you write them yourself?

Using AI to spin up new features is awesome for speed, but letting it handle the test suite usually creates a mess of brittle tests. They either pass by mocking out half the application or fail because the agent didn't understand the broader system architecture. Lately I've started treating AI-generated code as untrusted until I manually write or review the tests myself.

What kind of review workflow or safety checks do you use to make sure your test suite doesn't slowly rot away?