Nolan Lawson reports using multi-model agents on PRs to rank bugs by severity, yielding near-zero false positives while exposing pre-existing flaws that predate the changes under review (https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/). Lawson's workflow—Claude sub-agents, Codex, and Cursor Bugbot followed by manual validation—routinely identifies critical security, correctness, and performance issues, prompting abandonment of misguided approaches in some cases. This process aligns with patterns observed in earlier agent evaluations where repeated model passes increased defect detection rates without proportional speed gains.