The Claude Code team doubled merged pull requests while reverts fell over 40%. Their review rules fit client automations too

The question is not whether a human read the change. It is what could go wrong and which check should catch it.

What happened

The Claude Code team published how it reviews code now that agents write most of it. From June to September, merged pull requests to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40%. In Anthropic's monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September.

Their rules, in short. Claude has to verify its work with something it can run, like tests or a build, and show the output instead of saying it is done. Every pull request gets an AI review, and findings are checked by other agents before a human sees them. The change sets the risk tier, not the author: a classifier picks the tier, sensitive paths always need a person, and an agent's approval never counts as a person's. Whoever prompted the change is accountable for it.

They also found AI review loops can run away. One pull request went through more than fifty rounds of review and fixes overnight. Now the reviewer asks the author to step back after three rounds. Past incidents became an eval set the reviewer runs against daily, and feature flags can be turned off without a deploy. Simon Willison highlighted the core framing: "did a human read the code?" is an imprecise proxy for what could go wrong.

My take

Business automations change constantly: a new field, a new branch in a workflow, a prompt tweak. Most of those changes get no review at all. These rules scale down well:

1. Tier changes by what they touch. A wording change in an internal Slack alert is low risk. Anything that sends to customers, moves money or deletes CRM records gets a human, every time.
2. Make the builder agent show a real test run, with the input it used and what came out.
3. Turn every incident into a test case you rerun before the next change.
4. Give every automation an off switch that does not need a redeploy.
5. Cap AI fix loops. Three rounds, then a person looks at the approach.

Small teams do not need a merge queue. They need to decide in advance which changes are allowed to ship without a second look.