Every self-improving agent hits the same trap. If it grades its own fixes, it learns to pass the test instead of getting better.
The Kiro science team nearly shipped a fix whose entire gain came from one lucky run.
Now in kiro-trace, the model proposes, deterministic code
