Build Daily
Tinley Park · October 21, 2026Claude Writes 90% of Its Code. Here's How They Handle Review
This video breaks down how Anthropic's Claude Code team reviews code now that Claude writes roughly 90% of its own codebase, walking through six layers of automated checks, risk-scaled human approval, a 50-round review loop, and the new Claude Code Review product. It also covers Scion, the system that replaced git on Anthropic's monorepo, and what other teams can copy from the approach.
The outline
Tap a timestamp to jump there in the video.
Intro
The video introduces the topic: Claude now writes about 90% of Claude Code's own code, merged PRs more than doubled, and reverts fell by over 40%.
Why Code Review Broke
It explains why traditional code review broke down when an AI agent generates the vast majority of pull requests, overwhelming human reviewers.
The Goal: One Human in the Loop
The team's guiding principle is to keep exactly one human in the loop rather than trying to review every line of AI-generated code.
Layers 1–4: Before a Human Looks
The first four layers of the review pipeline are fully automated checks that run before any human ever looks at the code.
Sponsor: OpenRouter Armor
A sponsor segment promotes OpenRouter's new Armor product for securing AI applications.
Layer 5: Human Approval by Risk
Layer 5 introduces human approval that is scaled by risk, so low-risk changes need minimal oversight while high-risk changes get deeper scrutiny.
Layer 6 and Fast Rollback
Layer 6 and the fast rollback mechanism ensure that if something does slip through, it can be reverted quickly and safely.
The 50-Round Review Loop
The team discovered a 50-round review loop in which Claude iterates on its own code, and they learned to cap the number of rounds to avoid infinite cycles.
Where Claude Code Review Fits
The new Claude Code Review product is positioned as an advisory tool that reads REVIEW.md and can be triggered via @claude review or /code-review, but it never blocks a merge.
Scion: They Replaced Git Too
Anthropic replaced git on its monorepo with a system called Scion, and the video gives a quick look at what that entails.
What You Can Copy
The video distills practical takeaways that other engineering teams can copy from Anthropic's six-layer, risk-scaled review approach.
Outro
The outro wraps up the video and points viewers to the course, community, and newsletter links.
The 40% drop in reverted code matters more than the doubling of merged pull requests between June and September. For a solo builder shipping AI-generated code, the question stops being how fast the code is written and becomes whether the review loop catches what is wrong before it reaches production. Six layers of automated checks and risk-scaled human approval are a team-scale answer; the solo version is a shorter gate that still separates 'generated' from 'shipped.'