Kiro built an internal tracing system to find and fix hidden inefficiencies in its AI coding agent. The system proposes changes and validates them through deterministic code-based checks. It prevents false wins caused by outliers or luck. In its first week, it opened eight pull requests, with seven merged. The fixes targeted misleading tool descriptions, redundant file reads, and strict tool contracts. One fix removed a false claim seen 61,742 times. Another cut redundant file re-reads by 83%. A third cleared 205 schema rejections and 167 orchestration rejections. A silent file-search bug breaking one in eight searches was also fixed. A stricter reference agent revealed a relative-path bug affecting 13–17% of its reads but only 1.6% of Kiro’s. The loop improved compaction retention from 0.82 to 0.92. A smaller tool-and-prompt profile scored 82.4% pass@1 versus 77.5% on Terminal-Bench 2.1 at a third of the token cost. The team is reviewing whether to make this profile the default.
What changed
Kiro’s self-improvement loop graded its own fixes and risked shipping narrow hacks without regression checks.
What it unlocks
Automatically finding, diagnosing, and validating fixes for hidden inefficiencies and failures in Kiro’s agent behavior without human intervention.
- 61,742 exposures of false claim
- 83% reduction in redundant file re-reads
- 205 read-tool schema rejections cleared
- 167 sub-agent orchestration rejections cleared
What you need to act on it
- internal tracing system (kiro-trace)
- deterministic code-based acceptance gates
- separate roles for agent, proposer, and gate
Sources