Anthropic replaced human code review with AI to handle 30,000 PRs a day
Claude now writes 90 percent of its own code, forcing Anthropic to build a 7-layer automated review system just to keep up.
Tools · Source: Hacker News
What happened
Anthropic hit a massive scaling wall. Claude now writes about 90 percent of the code for Claude Code. That automation pushed their daily changes from 2,500 in May to over 30,000 by September. Manual code review broke down completely. Human engineers simply could not read that much code without slowing deployment to a crawl. They realized asking if a human read the code was the wrong metric.
They built a seven-layer automated review system to replace the manual bottleneck. Review no longer happens as a gate at the end of a pull request. It runs continuously like spellcheck while the code is written. Claude verifies its own work, runs tests, and reviews diffs before a pull request is even opened. Once opened, multiple AI agents hunt for bugs in parallel. A separate group of agents votes on each finding. Only verified bugs get flagged, while style nits are dropped entirely.
Humans only step in based on strict risk classifiers. A machine grades every change to determine the required human attention. Low risk updates get zero human reviewers. Medium risk gets one person. High risk gets two people. The results speak for themselves. Between June and September, merged pull requests doubled. At the same time, the share of code that had to be reverted dropped by more than 40 percent.
Key facts
- 90 percent — Share of Claude Code written by Claude
- 30,000 — Changes landing per day in Anthropic's monorepo by September
- 40 percent — Drop in the share of pull requests that had to be reverted
- 54% — Share of internal pull requests getting substantive review comments after AI review
Why it matters
This changes exactly how engineering teams must scale in the AI era. The old bottleneck in software was developer output. AI agents completely solved that problem. The new bottleneck is code review and quality assurance. Anthropic proves you can solve this by treating human attention as your most scarce and expensive resource. You must stop paying senior engineers to catch typos or syntax errors. You start paying them to manage system risk, handle complex architecture, and fix the rare things the machines miss.
The second-order effect is the complete death of the traditional pull request workflow. Review is shifting entirely to the left. If an AI agent writes the code, another AI agent must verify it before a human ever sees a notification. Teams will need to maintain explicit instruction files for their agents. You will need to tell the AI exactly what failure looks like in your specific codebase, or you will drown in false positives.
For builders
Automate your pre-commit verification
Do not let agents open pull requests blindly. Force them to run tests, build the app, or compare UI screenshots first. Anthropic estimates this verification step alone makes the final code two to three times better.
Write explicit review guidelines
Create explicit review documentation for your repository. Tell the AI exactly what failure looks like for that specific code. Cap the style nits and force the AI to cite specific file lines instead of guessing.
Implement risk-based human approvals
Stop requiring human review for documentation or internal flags. Use an automated classifier to route high-risk changes to senior engineers. Remember that the person who prompted the AI agent is always on the hook for the final code.
My take
I have said this before, but you cannot scale AI code generation without AI code review. If you are still forcing your senior engineers to read every single line of agent-generated boilerplate, you are burning cash and ruining morale. Build systems that catch bugs automatically. Save your humans for the hard architectural choices and incident recovery.
Original reporting: Hacker News. This is my rewrite and opinion.