Claude Code 团队如何做代码审查
Key highlights
Anthropic's team describes on the Claude.dev blog how Claude Code reviews code: machines handle most of the reading while human attention focuses on high-risk decisions, reshaping the old model of humans reviewing humans into a process of machine first review and human final review.
What happened
The numbers are convincing: from June to September 2026, merged pull requests more than doubled while the revert rate dropped over 40 percent; the monorepo's daily changes rose from about 2,500 lines in May to over 30,000 in September. In other words, AI did not make code chaotic, it let review capacity keep pace with explosive submission volume, proving the value of redesigning the process rather than simply adding reviewers.
Technical details
The core idea is splitting reading from judging: the AI reads hundreds or thousands of lines of diff, flags suspicious spots, and drafts initial comments, while human reviewers only make the final call on the high-risk areas the AI surfaced. That absorbs the volume surge while avoiding fatigue-driven misses under a brute-force manual process, putting scarce human attention where it is worth the most.
Versus the competition
Against the traditional everyone-reviews-everyone model, this division shines when AI produces code at high density. It does not miss architecture-level issues the way naive automated review can, and it scales with change floods better than pure manual review, a pragmatic middle ground between full automation and full manual effort that teams can adopt incrementally.
Industry impact and use cases
Put simply, when the main author of code becomes AI, the review process itself must be redesigned. For teams deploying coding agents at scale, this share is the most direct reference: keep humans on the most critical decisions rather than drowning them in diffs, so delivery stays both fast and stable.
Practical notes
For teams deploying agents, the most useful step is to define up front which decisions must be human-only, such as security and architecture changes, and then let the agent take over the remaining diff triage rather than trying to replace all review with a machine. Track revert rate and review latency as ongoing metrics, and as the agent's suggestions are stably accepted, widen its scope gradually while keeping the human gate firmly on architecture and security. This division, where machines absorb repetitive labor and humans guard the critical judgments, is the key to maintaining quality even after scaling.
Broader context
As the main author of code gradually becomes the agent, the center of gravity in the review process must shift as well. The traditional model of everyone reviewing everyone fails quickly when submission volume explodes, because human attention is a scarce resource that cannot be solved by simply adding more people. Splitting reading from judging, letting machines absorb the repetitive first pass, and keeping humans only on the highest risk decisions is a far more sustainable division of labor. This idea applies not only to code but to any knowledge work scenario where agents produce output at high frequency and at scale.
What to watch next
What deserves attention is how to define which decisions must remain human only. If the boundary is unclear, an agent may quietly overstep, or humans may be flooded by low risk suggestions and let their guard down. The safest approach is to list hard gates such as security and architecture first, let the agent take over the remaining triage, and track revert rate and review latency as long term metrics, widening permissions only after suggestions are stably accepted rather than letting go completely from the start.
Practical takeaway
For teams deploying agents, the most useful step is to define up front which decisions must be human only, such as security and architecture changes, and then let the agent take over the remaining triage rather than trying to replace all review with a machine. Track revert rate and review latency as ongoing metrics, and as the agent suggestions are stably accepted, widen its scope gradually while keeping the human gate firmly on architecture and security. This division, where machines absorb repetitive labor and humans guard the critical judgments, is the key to maintaining quality after scaling, and it is the part every organization should design early when bringing agents into the workflow.