Anthropic says Claude Code merges doubled, reverts fell more than 40%

3 persons at Anthropic described how the Claude Code team looks at code that AI agents write. From June to September, merged pull requests more than doubled. A pull request is a code change that is prepared to be added to the primary code. The share of changes with a revert fell by more than 40%. A revert is a change back to the code before the change.

How it works

Before a pull request opens, Claude does tests. A different agent then looks at the change in a clean context. Claude reviewers look for bugs, and other agents check the findings. A check for errors can stop a merge. Approval by a human is by risk level:

• Low risk: no approval, for example for documentation changes.
• Moderate risk: 1 person.
• High risk: 2 persons. Changes to CI workflows always have a human.

The numbers

The checks stop after 3 cycles. The share of pull requests with 5 or more cycles went from 8.9% to 1.5%.

Not known

These are numbers from Anthropic for its code. No external test is given. The source gives no number for incidents in live use.

Insight: Risk rules now decide what humans read

Anthropic says it merges more than 30,000 code changes each day, up from about 2,500 earlier in 2026. At that rate, humans cannot read each line. This reads as a move of human checks from the code to the rules that divide changes by risk. The persons who write those rules have the control. Know who sets your risk levels.