Anthropic’s Claude Code team says most code review can be automated

Anthropic’s Claude Code team says it uses automated checks and verification agents to handle much of its code review, while reserving human approval for higher-risk changes. The practical idea is not to remove review, but to spend people’s attention where the consequences are larger.

Anthropic Watch analysis

What happened

In a post published on 15 October, the team describes a review process in which most changes go through automated bug and security checks. Pull requests are sorted into risk tiers, which help determine whether human approval is needed. The team says it aims to reduce human review time without increasing software reverts or incidents.

It also describes using AI-generated explainers and open communication channels to help people keep track of the codebase while automated tools do more of the reading. Read the Claude Code team’s explanation.

Why it matters

AI coding tools can produce changes faster than people can carefully inspect them. Anthropic’s approach offers one answer to the resulting bottleneck: automate routine checks, apply more scrutiny to higher-risk work, and help developers maintain a shared understanding of what is changing.

That is a description of the team’s process, not evidence that the same approach will work equally well elsewhere. But it gives software teams a more concrete model than simply asking reviewers to read faster. The interesting question is where the risk tiers put the human sign-off.

Our read

The sensible part is treating review as a question of risk, not an all-or-nothing choice between a person checking every line and nobody checking anything. Automated checks can clear routine work, but the team still needs to know what changed and which changes deserve a closer look. No magic wand, then, just a rather consequential system of triage.

What to watch

• How Anthropic defines the risk tiers and decides when human approval is required.
• Whether the team publishes measured results for reverts, incidents and review time.
• How well AI-generated explainers help developers keep their bearings as code changes faster.

Discussion spark: Should software teams require human approval for every AI-assisted code change, or reserve it for changes their risk checks flag as consequential?

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