AWS's kiro-trace mines Kiro's own runs for fixes, but only deterministic checks decide what ships
The internal loop finds where Kiro fails or wastes effort, proposes one prompt or tool change at a time, and keeps it only if a code-based gate (including a full-benchmark run against a re-run control) and then a human reviewer accept it; in its first week it opened eight PRs and seven merged, including an ~83% cut in redundant file re-reads and a compaction fix that raised retention from 0.82 to 0.92. A leaner tool-and-prompt profile that scored 82.4% vs 77.5% pass@1 on Terminal-Bench 2.1 at about a third of the token cost is still under review, and all figures are Amazon's internal experiment deltas.
