How the Claude Code team reviews code when Claude writes most of it

Addy Osmani's article, How the Claude Code team does code review (opens in a new tab), describes a setup where Claude writes about 90% of Claude Code's code. From June to September, merged PRs more than doubled and the share of reverted PRs fell by over 40%.

The aim is for most PRs to merge with at most one human in the loop, without more reverts or incidents. Human time counts as a cost, including time spent fixing bad code.

A few things I took from it:

• Layers, then rollback. Pre-PR verification, AI review on every PR, a security review that can block a merge, CI checks, human approval scaled to risk, then a merge queue with feature flags and staged rollouts. Fast revert is the last safeguard.
• Approval depends on the change, not the author. Agent approvals don't count as human ones, and an approval covers a single commit, so new pushes need a fresh look.
• Cap the review loop. One PR bounced between an agent author and an agent reviewer more than fifty times overnight. The team now limits it to about three cycles.
• Explainers instead of diff-reading. Engineers ask Claude to explain the part of the code they're about to change.
• "You land it, you own it." Whoever fixes an issue owns it, and incident reviews look at the system rather than the person.