Story · Kiro (AWS)
How Kiro fixes itself, and how we keep the fixes honest (Kiro (AWS))
blog post · Story page
One false sentence in Kiro's tool descriptions reached the model 61,742 times across a few hundred sessions. Judging fixes is harder: one that passed more tasks on fewer tokens owed almost all its gain, once split by host, to a single run where the agent escaped a loop it usually gets stuck in.
In plain words
- Kiro nearly released a change that looked helpful because of an unusually successful test run.
- During that run, the artificial intelligence assistant escaped a loop of repeated actions that usually kept it stuck.
- Grouping results by test environment showed that this unusual run explained almost all the apparent improvement.
- Without that environment's results, the change was slower, more expensive, and less accurate.
- For developers, a higher overall score can hide a change that performs worse elsewhere.
Appeared in
- Most of a harness-evolution gain is overfitting and extra compute
Oct 22, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.