Anthropic's AI code review: real-issue detection jumped from 16% to 54%
dotey · x · 2026-11-20
Claude Code PM Morgan Lunt shares how Anthropic overhauled its code review process.
- A year ago every PR was human-reviewed; once models could finish features solo, PR volume exploded and review became the bottleneck (he cites a Harvard study: PR review wait times rose 56% after adopting AI coding agents). Internally, engineers rubber-stamped each other's PRs — he calls it a "kabuki theater" phase.
- Homegrown review bots proliferated: up to seven bots per PR with conflicting demands, causing 50–200 rounds of rewrites that drifted from the author's intent, because the bots lacked shared context and goals.
- Fix: consolidate all review concerns into one master review agent that deduplicates internally; humans sign off only on high-risk changes, targeting 90% of PRs merged with no reviewer other than the author.
- Results: PRs with substantive issues caught rose from 16% to 54%, merged PR count doubled, rollbacks dropped 40%. Lunt likens it to self-driving: it needn't be perfect, just better than humans.
More from coding & agent
- Cognition: Gemini 4 Neon scores 46.2% on FrontierCode 2.0, near Opus 5.5 at 1/3 cost — scaling01 · 2026-11-20
- HKU Open-Sources PhoneGo, a CLI+GUI Hybrid Phone Agent Cutting Steps by 47% and Latency by 56% — aigclink · 2026-11-20
- Modular rebuilds MAX's Module API with PyTorch-like syntax and eager mode — clattner_llvm · 2026-11-20
- Apple reportedly building an AI agent that controls devices, browses the web, and codes — Polymarket · 2026-11-20
- comfy-ts: open-source typed ComfyUI workflows for humans and coding agents — rvion · 2026-11-20
- Feedbacker Adds MR-Style Inline Code Review to Claude Code and MCP Agents — JeremyCMorgan · 2026-11-20