How Anthropic reviews code when Claude writes 90% of Claude Code
AGI Hunt · wechat · 2026-10-16
Anthropic details how the Claude Code team reviews code when Claude writes 90% of it: shift verification and review earlier, and scale human approval by risk. From June to September, merged PRs doubled while rollback rate dropped over 40%.
- Layered checks: Claude must demonstrate verification (tests, builds, screenshots) pre-PR, a fresh-context subagent reviews the diff, then Claude code review (substantive comments rose from 16% to 54%) and a blocking security review
- Risk-based approvals: docs can merge with zero human approvals; sensitive paths and high-risk code require 1-2 humans; agent approvals never count as human, and whoever directs the agent is accountable
- Taming review noise: one CLI PR went through 50+ overnight review rounds; git blame showed 59% of repeatedly flagged issues existed at first review. Now: 3-round cap, re-review only new diffs, fix whole bug classes. Single-pass merges rose from 65% to 78%, average review rounds fell from 2.08 to 1.32
- Every incident becomes an eval case; the reviewer runs daily regression evals. Humans keep direction calls, architecture, high-risk code, and ownership
Related event: Claude Now Writes 90% of Claude Code, Anthropic Reveals Review Process(2 posts)→
More from coding & agent
- AI coding cost per output hour falls 15% in September after 4x climb — AccBalanced · 2026-10-17
- Claude Code 2.1.302: prompt tokens jump 39% in a day, system share rises — ClaudeCodeLog · 2026-10-17
- Full changelog: Claude Code 2.1.302 adds auto-mode model env var — ClaudeCodeLog · 2026-10-17
- Claude Code 2.1.302 ships 63 CLI changes, blocks 128MB+ image loads — ClaudeCodeLog · 2026-10-17
- Agent ports a Rails app to Jai with one prompt: 'things move fast' — lucasmeijer · 2026-10-17
- From 'git push heroku master' to telling Codex: keep my app deployed and healthy — alex_frantic · 2026-10-17