The bounded-loop framing is the right one. Most agent failures in production aren't model failures — they're loop failures. No clear stop condition, no proof that state changed, retries that keep going after the useful signal is gone. Once you force the loop to justify another turn, most of the fake progress disappears.