Anthropic discloses Claude Code Approximately 90%of the team's code is written by Claude, and it has redesigned its AI code-review process

AT A GLANCE

The report stated that Claude Code approximately 90%of the code is written by Claude, with multi-Agent stage reviews and tiered human approval.

Article

Claude Code Begin allowing AI reviews AI：90%Code written by Claude; low-risk changes require no human approval

Beating Insights AI flash news: Anthropic disclosed that Claude Code approximately 90%of its code has already been written by Claude. This year,June to September，Claude Code the number of code change requests merged by the team (PR) more than doubled, while the proportion requiring rollback fell by more than 40%.

To keep pace with the speed at which AI writes code, Anthropic redesigned its code-review process. Before each submission, Claude runs tests and has independent sub-Agent inspect the changes. After submission, multiple Agent search for Bug in parallel, followed by another group of Agent conducting cross-validation. Security reviews and automated testing also prevent problematic code from being merged.

Human review is tiered by risk. Low-risk changes, such as documentation edits submitted by trusted developers, may be merged directly after passing automated checks; ordinary product changes require approval from 1 people, while high-risk code requires 2 people.AI's approval cannot replace humans; developers remain responsible for the final code.

However,AI reviews can also fall into an infinite loop. Anthropic disclosed that one Claude Code change underwent more than 50 rounds of review and fixes overnight. The team's examination found that approximately 59%of issues involved code that had already existed during the first review round, rather than new issues introduced by subsequent changes. The team then adjusted the rules: after more than 3 review rounds, the author must reconsider the approach, and repetitive feedback is limited to prevent Agent from becoming entangled with one another.

Original link https://m.theblockbeats.info/flash/372426

Key points

The article states that Anthropic disclosed that Claude Code approximately 90%of its code has already been written by Claude.
The article states that, this year,June to September，Claude Code the number of PR merged by the team more than doubled, while the proportion requiring rollback fell by more than 40%。
Before each submission, Claude runs tests and has independent sub-Agent inspect the changes.
After submission, multiple Agent search for Bug in parallel, followed by another group of Agent conducting cross-validation.
Low-risk changes, such as documentation edits submitted by trusted developers, may be merged directly after passing automated checks.
Ordinary product changes require approval from 1 people, while high-risk code requires approval from 2 people.
AI's approval cannot replace humans; developers remain responsible for the final code.
One Claude Code change underwent more than 50 rounds of review and fixes.
The team's examination found that approximately 59%of issues involved code that had already existed during the first review round.
The team subsequently stipulated that after more than 3 review rounds, the author must reconsider the approach, and repetitive feedback is limited.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This process enables AI not only to write code, but also to participate in testing, identifying Bug and cross-validation; whether humans intervene and how many people must approve depend on the risk of the change.

Why it matters to readers

The article presents a review approach for handling AI rapid code output, while also explaining that automated reviews can enter repeated loops and still require rule-based constraints and human accountability.

Beginners should note:AI generating code does not mean testing, review, and confirmation of responsibility can be skipped; the higher the risk of a change, the stricter the human-approval requirements.

Risks and unknowns

The material is a news brief summary and does not include Anthropic's original disclosure materials.
Approximately 90%, more than doubled growth, a decline of more than 40%and approximately 59%and other statistical measures cannot be confirmed from this article alone.
The article does not explain code quality, the number of review rounds, or the proportion of human involvement after the rules were adjusted.

Anthropic discloses Claude Code Approximately 90%of the team's code is written by Claude, and it has redesigned its AI code-review process

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center