Self-improvement becomes operationally credible only when proposal and evaluation are separated by checks the proposer cannot control.
A constrained change loop in which run evidence generates candidate edits, while an external-to-the-model gate and human review control whether those edits ship. Candidates face staged checks, including regression checks and a full-benchmark comparison against an unchanged control. Amazon reports that the system opened eight pull requests during its first week; seven merged and one was closed for unrelated reasons. Agent behavior can fail in tool sequencing, error recovery, or long-context work without producing a neat bug report, making manual inspection of thousands of runs impractical.
Teams building agents should treat evaluation architecture as part of the product: improvements need a control, regression coverage, and an accountable merge boundary, not just a model-generated patch. All figures and evaluations described here come from Kiro’s internal Amazon usage and experiments; they do not establish a production-wide effect or generalize automatically to other agents. Teams considering automated agent improvement should invest in independent acceptance tests before increasing the autonomy of the change generator. Whether future reports disclose durable results across broader task sets and whether rejected candidates or regressions are quantified.