The company says the system uses staged checks, runs a full-benchmark comparison before a change proceeds, and relies on deterministic code—not the model that proposed a change—to accept or reject candidates. Changes that pass become pull requests for human review. In its first week, kiro-trace opened eight pull requests: seven merged and one closed for unrelated reasons. The reported results concern Amazon’s internal usage and experiments, not production-wide estimates. Source: https://kiro.dev/blog/how-kiro-fixes-itself
On Tuesday, Amazon unveiled Kiro-trace , an internal tracing system that improves Kiro, Amazon's AI coding agent, from its own runs.
In one minute
Amazon has described kiro-trace, an internal system that uses Kiro’s runs to identify failures or wasted effort, propose changes to prompts or tools, and test those changes against an unchanged control.
What changed: Amazon’s internal kiro-trace system analyzes Kiro’s runs, proposes prompt/tool changes, and uses deterministic checks to accept them; it opened eight pull requests in its first week.