Claude Code team: 90% of code AI-written, reverts down 40% with automated review
addyosmani · x · 2026-10-26
Anthropic's Claude Code team details how it does code review when Claude writes 90% of the code: PRs merged doubled between June and September while revert share fell over 40%.
Key practices:
- Lifecycle of a change: pre-PR verification by Claude plus a second agent reviewing the diff in fresh context; required CI; parallel agent bug hunts with verifier voting; a separate security review
- Risk-based approval: a classifier sets how many humans (0, 1, or 2) approve; some paths always need a person
- After merge: flag-gated internal-first rollout, gradual release, fast rollback
- Incidents are the eval: every incident/revert becomes an expert-checked eval case; reviewer changes are kept only if daily eval scores improve
- Layers, not one gate: six layers of checks with fast rollback underneath
A complete blueprint for scaling code review when agents write most of the code.
More from coding & agent
- Open-source sky camera run by Claude Code catches 17 meteors overnight, cuts star false alarms to 7% — Terrible_Award_2124 · 2026-10-26
- AI-built OpenSearch clone validated by 6,000 black-box tests: Rust at 64.5%, Go at 32.6% — Strong-Mousse-1020 · 2026-10-26
- Fully local open-weight pipeline produces 10-min animated episodes at ~4 GPU-hours each — Zangetsu__0 · 2026-10-26
- Even at max Daybreak tier, AI vuln-analysis requests still blocked by cybersecurity guardrails — jd_pressman · 2026-10-26
- Bambi quietly gains nonlinear model support, with a blog post showing how — twiecki · 2026-10-26
- Dev used Claude Opus 5.5 agents for 5 days to compose and render a 3D pianist cinematic — 141_1337 · 2026-10-26