Anthropic’s Claude Code team has stopped trying to have a person read every change to its own product, and the rule it replaced that with is the part worth copying. How many humans must approve a change depends on how risky the change is, and a classifier, not the author, makes that call.
The team’s own account, published on the Claude Code blog, starts from a figure it has given before: Claude writes about 90 percent of Claude Code’s code. Review does not scale the way generation does. In Anthropic’s separate monorepo (the company’s main shared codebase), changes landing per day rose from about 2,500 in May to more than 30,000 in September.
The ladder
There are three tiers. Low risk covers documentation, copy, internal-only features, and additive changes behind a feature flag (a switch that keeps new behavior off until someone turns it on). Those need no human approval once the automated review and security checks pass. One person must approve the middle tier: ordinary product changes, along with delicate areas like system prompts, the auto-updater, and caching. The top tier, for the riskiest code plus a few named paths that never skip it, needs two.
The post is explicit about who decides. “The change sets the tier, not the author.” A classifier looks at each change and assigns the tier. A second one checks for the exception: a change whose tier would not call for another person, but which should get one anyway. Some paths are hard-wired to require a human regardless: any file that defines the automated test workflow itself, and anything the code owners have marked sensitive.
That is the load-bearing decision, and the obvious weak point. If the classifier files a risky change as low, the ladder routes it past every human by design. The post does not report how often the classifier is wrong, or how it is tested. What it does show is the hedging around it: fixed lists of paths that always need a person, lighter routes only for authors on a trusted list, and approvals tied to one specific commit, so new commits the system cannot vouch for force a fresh sign-off. A bot’s approval never counts as a human’s. And whoever directed an agent to write the change is treated as a required reviewer.
Review while writing, not after
The second idea is to move review earlier. The post says review now runs “continuously, more like spellcheck while you write,” so a human is usually asked to look only after Claude has already reviewed the change and repaired what it found. Engineers give Claude tests or a design screenshot to check against and ask for the output, not a claim of being done. Then a separate agent with fresh context reviews the diff.
The team also found a failure particular to agent authors. An agent writing code and an agent reviewing it can generate work for each other without end. In one case an overnight loop of critique and repair ran past fifty cycles before the author stopped it. About 59 percent of what the reviewer flagged, by the team’s git blame check, was code present since its first pass. The fix was a cap of three rounds, then review of only the new changes.
Where the humans went
People were not removed. The post says the shift is away from reading each diff and toward five decisions: owning the outcome, judging the evidence the checks produced, whether to build something at all, risky code like billing and the auto-updater, and architecture that sets a pattern for others. The stated goal: most changes merge with at most one person involved, who can be the author.
The numbers, and their limits
There are some. From June to September, merged changes more than doubled, and the share reverted dropped by over 40 percent. The share of internal changes receiving substantive review comments rose from 16 percent to 54 percent after the review tool launched. Those are the only outcome figures. There is no count of how many changes clear without a person, no latency data, and no tally of defects caught. A falling revert share is also not the same as fewer incidents.
This is Anthropic’s team describing Anthropic’s tooling on Anthropic’s blog, and the product it uses is the one the company sells. The account is self-reported and unaudited. The post concedes a gap: no circuit breakers yet that tighten approval when revert or incident numbers rise, and engineers there want them in place before further code merges with no human involved.
Based on the Claude Code blog post “How the Claude Code team does code review” by Addy Osmani, Morgan Westlee Lunt, and Qing Wang, 15 October 2026
