Anthropic's Claude Code team has published how it reviews code when agents write nearly all of it. Claude authors about 90 percent of the Claude Code codebase, and the team says merged pull requests more than doubled between June and September while the share that had to be reverted fell by more than 40 percent. Across Anthropic's monorepo, changes landing per day went from roughly 2,500 in May to more than 30,000 in September.
The workflow pushes most of the reading to machines. Before a pull request opens, Claude runs the tests, build and linter, and a second agent reviews the diff in a fresh context. On every PR, required CI gates the merge, parallel Claude reviewers hunt for bugs with verifier agents voting on each finding, and a separate security pass blocks anything left open. A risk classifier then sets the approval tier: no human, one, or two. Some file paths always require a person.
For founders building with agents, the lesson is that review is becoming continuous rather than a gate at the end of the road. The team tracks human approvals per PR, human comments and how long a change waits on a person, with revert and incident rates as the guardrail: if those rise, review time was moved rather than saved. Their rule for authorship is blunt, you land it, you own it, backed by feature flags and gradual releases so mistakes can be fixed forward or rolled back fast.