Claude writes 90% of Claude Code, so its team rebuilt code review around risk tiers

Seven layers, and one rule about who may approve

The post says Claude writes about 90% of Claude Code's own code. From June to September merged pull requests more than doubled, and the share that had to be reverted fell by more than 40%. The team kept code review but moved most of the reading to automated checks. The goal is most PRs merging with at most one human involved, tracked by human approvals per PR, human comments and wait time, with revert and incident rates as the guardrail.

The layers: Claude runs tests, builds and linters and a fresh-context subagent reviews the diff before the PR exists, guided by CLAUDE.md and a code-owner-written REVIEW.md; parallel review agents hunt for bugs while verifier agents vote on each finding, with only agreed findings posted as Important, Nit or Pre-existing; a separate security pass can block merging; required CI; a risk classifier sets low (no human), medium (one person) or high (two people); then a merge queue, feature flags that ship to employees first, and fast rollback.

The numbers on review convergence: PRs passing after one review-and-fix round rose from 65% to 78%, PRs needing five or more rounds fell from 8.9% to 1.5%, mean reviews to pass dropped from 2.08 to 1.32. About 59% of what the reviewer kept flagging was old code present since its first pass. Humans still own whether to build something, risky code such as billing and the auto-updater, architecture, and accountability: "You land it, you own it."

The thread: throughput is not knowledge

The first reaction was ridicule of the scale itself.

"I really don't get the tendency here from Anthropic to describe what's most likely churn?" — a commenter skeptical of the PR volume

That reply came from the co-founder and CTO of Goody, who pointed out the sums were being misread:

"Claude Code is in a separate repo. So it's not the TUI that has 1,000 PRs a day." — Mark Bao, co-founder and CTO at Goody

A nine-year full-stack engineer was unmoved by the correction:

"When you get to the point that you are doing 1000 PRs I have to seriously wonder what even are you building" — a full-stack engineer with nine years in the industry

The deeper objection was about understanding. When the post's idea of an LLM summarizing a slice of the system came up, one reader said the output is just search by another name, and another:

"There is no knowing the code without experiencing the code." — a commenter who reads tests to check what they verify

What to take from it

The most transferable detail is the failure mode, not the stack: author-agent and reviewer-agent loops do not converge on their own, and with 59% of repeat findings being old code, later rounds mostly find older bugs. A round cap and diff-only re-review are cheap to copy; the risk classifier is the part that needs trust in your own incident data.