Type: url
Original: https://techblog.openwork.co.jp/entry/ai-code-review-3-layer-architecture
Status: generated
Summary
An OpenWork engineer describes a three-layer agent orchestration built on GitHub Copilot CLI custom agents to raise AI code-review precision. A parent agent decides which technical domains a branch touches and fans out per-perspective review agents; each perspective is reviewed by four agents split across two model families, and only findings two or more agents raise independently are kept, which is the post’s stated defence against hallucinated and trivial comments. Results post as PR inline comments, and a context-collection agent feeds prior review threads back in so engineers’ “won’t fix” replies are not re-raised.