How the Claude Code team does code review
Points and comments are a snapshot, not live.
Claude writes 90% of Claude Code's code; automated review replaces most human line-by-line reading.
The Claude Code team redesigned code review for agent-written code. Merged PRs more than doubled from June to September while revert rates fell over 40%. Most review is now automated, with six layers: pre-PR verification, AI review on every pull request, security review, blocking automated checks, human approval scaled by risk, and a merge queue. Human time is reserved for decisions machines are bad at: whether to build something, risky code, architecture, reading evidence, and owning the outcome. A risk classifier sets the review tier. Incidents become eval data to improve the reviewer. The goal is most PRs merge with at most one human in the loop.
What commenters are saying
Skepticism dominates the thread. Top commenters question the 30,000 daily changes figure, calling it churn masquerading as productivity. Several point to Claude Code's GitHub issue tracker with over 5,000 open issues, many AI-generated and locked, as evidence of regressions. One commenter argues the approach moves fast only to break things, then fix them endlessly. Another defends it as roughly 10 PRs per engineer, comparable to human output scaled by agents. A separate thread warns against dismissing security review findings without human AppSec involvement.