Kiro@kirodotdev
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 decides, and a human merges.
The fixes that survived:
→ Redundant file re-reads down about 83%
→ Compaction retention up from 0.82 to 0.92
Full breakdown 👉
126435K












