Anthropic 如何在 Claude 写下 Claude Code 约 90% 代码后重新设计代码评审
Key highlights
Anthropic has shared how its Claude Code team re-architected code review after Claude began writing roughly 90% of the code, fundamentally changing how humans and machines divide the labor and shifting review away from the old "everyone reviews everyone" habit toward a "machine first pass, human final pass" model.
What happened
Per the author's reading of Anthropic's blog, the team now hands most of the reading to machines while reserving human attention for the highest-risk decisions. The numbers are striking and worth repeating: from June to September 2026, merged pull requests more than doubled while the revert rate dropped by over 40%. In other words, throughput went up and quality got steadier simultaneously, with none of the "more AI code means more chaos" that skeptics predicted when adoption began in earnest.
Technical details
The redesign splits review into two layers. The first layer has AI perform large-scale context reading, style checks, and flagging of potential bugs and risk points across sprawling diffs. The second layer has human reviewers touch only the high-uncertainty areas the machine surfaced. That structure absorbed the surge in change volume while avoiding the fatigue-driven misjudgments of a brute-force manual process. The monorepo's daily change count climbed from about 2,500 lines in May to over 30,000 in September, a level of throughput this workflow was explicitly built to sustain rather than to staff up for with more reviewers.
Versus the competition
Against the traditional "everyone reviews everyone" model, this "machine first pass plus human final pass" division fits AI-heavy teams far better. It does not miss architecture-level issues the way naive fully automated review can, and it scales with volume far better than pure manual review, occupying a pragmatic middle ground between full automation and full manual effort that most organizations can adopt incrementally without a risky big-bang change.
Industry impact and use cases
Put simply, once AI becomes the primary author of code, the review process itself has to be rewritten rather than force-fit into the old one. For any engineering team already deploying coding agents at scale, this retrospective is a usable playbook: keep humans on the most critical decisions and let machines absorb the repetitive triage, so the team moves fast without sacrificing stability.
Practical notes
The lesson transfers cleanly to other teams: define a small set of high-risk decision points that always require a human, and let the agent handle the bulk of diff triage in between. Instrument both revert rate and review latency as ongoing metrics; as the agent's suggestions are consistently accepted, gradually widen its scope while keeping the human gate firmly in place on anything touching architecture or security.
Broader context
The Anthropic example is worth reading as a template rather than an anecdote, because the underlying dynamic repeats wherever agents write code at scale. When the machine produces most of the diffs, the scarce resource becomes human judgment about architecture, security, and product trade-offs, not hours spent reading line-by-line. Organizations that cling to the old "everyone reviews everything" norm will either bottleneck their own velocity or quietly skip review and accumulate risk. The healthier pattern is to formalize which decisions are human-only, automate the rest, and treat the agent's acceptance rate as a leading indicator of where it can safely be given more autonomy. The same logic extends beyond code to any domain where an AI produces draft work that a human ratifies.
Broader context
The Anthropic example is worth reading as a template rather than an anecdote, because the underlying dynamic repeats wherever agents write code at scale. When the machine produces most of the diffs, the scarce resource becomes human judgment about architecture, security, and product trade-offs, not hours spent reading line by line. Organizations that cling to the old "everyone reviews everything" norm will either bottleneck their own velocity or quietly skip review and accumulate risk. The healthier pattern is to formalize which decisions are human-only, automate the rest, and treat the agent's acceptance rate as a leading indicator of where it can safely be given more autonomy. The same logic extends beyond code to any domain where an AI produces draft work that a human ratifies, from legal review to content moderation. The metric that matters most is not how fast code is produced but how rarely a bad change reaches production, and that is exactly what a two-layer review is engineered to protect.