Discussion

Anthropic’s six-layer review system shows how Claude Code handles a surge in AI-written code

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#11937

Anthropic says Claude now writes about 90% of Claude Code’s code, while the number of pull requests merged into the codebase more than doubled between June and September. The team’s answer is not to scrap review, but to automate much of it and reserve human approval for changes whose risks warrant it. That is a consequential shift in how software teams might manage AI-generated work, not merely a clever new coding trick.

Anthropic Watch analysis

What happened

A report by BigGo Finance describes the Claude Code team’s six-stage review process, which Anthropic disclosed on 15 October. It includes tests and builds before a pull request is opened, automated reviews and security checks, continuous integration, risk-based human sign-off, and a final rerun of checks before merging.

Low-risk changes can skip human approval after automated and security checks. Ordinary product changes need one approver, while the highest-risk changes need two. The team also found that automated reviews could get stuck repeatedly flagging old code. After changing the review rules, the share of pull requests passing after one round rose from 65% to 78%, while those needing five or more rounds fell from 8.9% to 1.5%, according to the report.

Why it matters

The practical problem is familiar to any team using coding agents: when code arrives faster, a line-by-line human review of every change can become the bottleneck. Anthropic’s approach shifts much of the checking to automated systems, while tying human approval to risk. That offers a useful, concrete example of how a team is adapting its workflow as AI-generated code scales up.

The safeguards are not a guarantee. The report describes a bug that passed multiple rounds of human and automated review, and says the team has not yet built a proposed circuit breaker that would tighten approval requirements when incident metrics worsen.

Our read

This is a more useful story than “AI wrote the code”: it shows what the team does to check that code, and where it still depends on people. Risk-based sign-off makes sense as a way to spend scarce reviewer attention, but the system’s success depends on its tests catching the right things and its escalation rules responding when they do not. The promising part is the use of past incidents as a test bank; the missing circuit breaker is the bit that deserves a raised eyebrow.

What to watch

  • Whether Anthropic builds the proposed circuit breaker and explains what metrics trigger it.
  • Whether the review test bank keeps detecting bugs as models, prompts and tools change.
  • Whether the risk-based approval model spreads beyond Claude Code, and what evidence teams publish about its results.

Discussion spark: Should teams let low-risk AI-written changes merge without human approval after automated checks, or should a person sign off every change?

Sources and evidence

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.