How Anthropic reviews code when Claude writes 90% of it

When Claude writes roughly 90% of the code, how does Anthropic keep quality in check? Addy Osmani of the Claude Code team offered a systematic answer in a long thread on October 29. The backdrop is a dramatic shift in scale: daily merges in Anthropic's monorepo surged from about 2,500 in May to over 30,000 per day by September—no human can read every diff, so the traditional human review model had to be rebuilt.

Confirmed

• Ownership is clearly defined: Osmani stressed "you merge it, you own it"—code still has owners in the AI era; bot approvals never count as human approval, and PRs opened by agents are signed off by the person who ran the agent.
• Tiered reviews: Review depth is determined by the risk of the change, not by who authored it—typo fixes and automated updates go through lighter review paths.
• Multi-agent debate reviews: Multiple agents hunt for bugs in parallel, debate each other, and use voting to weed out false positives, while filtering out nitpicky style comments; every finding is either fixed or closed with a rationale. The share of PRs receiving substantive review comments rose from 16% to 54%.
• Review-first completion: Claude can't just claim it's "done"—it must first pass tests and builds, compare the UI against design specs, and prove completion with evidence; then a fresh agent (with no context contamination) reads the entire diff to review. Per the Claude Code author, this lifts quality by 2-3x.
• The last line of defense is undo: When AI writes most of the code, the final safeguard isn't a reviewer but feature flags, staged rollouts, and one-click rollback; every incident feeds back into improving the review process.

Why it matters

These practices answer the industry-wide anxiety of "how do humans gatekeep when AI writes code at scale": the answer isn't making humans read more diffs, but splitting review into four layers—machines proving completeness, multi-agent cross-debate, humans owning merge decisions, and fast rollback as a safety net. The jump in substantive review comments from 16% to 54% shows the process redesign turned review from a formality back into something genuinely effective—a directly useful reference for engineering teams adopting AI coding tools.

2026-10-29 ~ 2026-10-29 · 6 related posts

• [source] How Anthropic Reviews Code When Claude Writes 90% of It — ccerrato147 · 2026-10-29
• Anthropic's monorepo went from 2,500 to 30,000 merges/day with Claude writing ~90% of code — ccerrato147 · 2026-10-29
• Review moved before the PR: Claude must prove completion with tests; Boris Cherny cites 2-3x quality gain — ccerrato147 · 2026-10-29
• [source] Anthropic runs parallel bug-hunting agents with voting; real review comments rose from 16% to 54% — ccerrato147 · 2026-10-29
• Code ownership in the AI era: you land it, you own it, bot approvals never count as human ones — ccerrato147 · 2026-10-29
• [source] Anthropic's review philosophy: undo, not reviewers, is the last line of defense for AI-written code — ccerrato147 · 2026-10-29