
Must read
๐ค Amazon's kiro-trace: a coding agent that fixes itself, with a guard
Kiro-trace reads failed agent runs, diagnoses the cause, proposes one change at a time, and keeps it only if deterministic code passes it. A human then reviews it. In week one it opened eight PRs and seven merged. Fixes include a wrong tool description that caused 61,742 errors, and a search bug breaking about 1 in 8 searches. Long-task retention rose from 0.82 to 0.92.
Why it matters: It is a concrete pattern for improving your own agent harnesses without letting the model grade itself.
Takeaway: Keep the proposer separate from the judge, re-run a control each time, and turn each fix into a permanent regression test. Try removing rules before adding them.
Read it fully? Yes, if you build agents. Numbers are internal.







