Claude Code’s review redesign: scarce resource is human attention; guardrail is reverts

On 15 October 2026, Claude.dev published an engineering essay by Addy Osmani, Morgan Westlee Lunt and Qing Wang that refuses to hide the tension: the team says Claude writes about 90 percent of Claude Code’s own code; from June to September, merged pull requests into that codebase more than doubled, while the share that had to be reverted fell by more than 40 percent. The live question is therefore not whether anyone still reads code, but what review is for when agents push change volume past line-by-line human reading.

Their answer is blunt. They did not remove review. They moved most of the reading to machines, defined the scarce resource as human time and attention, and invested heavily in making mistakes fast to undo. That is a company self-report, not a third-party audit—but it states an operational claim that later revert and incident numbers can falsify.

The piece splits code review into two jobs: find and fix issues before they ship, and build shared context across the team. The article focuses on the first. Anthropic’s monorepo went from about 2,500 landings a day in May to more than 30,000 in September; Claude Code sits in a separate repository and still feels similar pressure. “Did a human read the code?” is treated as an imprecise proxy. The sharper question is what could go wrong with this change, and which check—automated or human—is the right one to catch it.

The objective function is rewritten too: most pull requests in the Claude Code repos (CLI, desktop and web, mobile) should merge with at most one human in the loop—who may be the author—without raising reverts or incidents. They measure human approvals per PR (median and 95th percentile), human comments, and how long a PR waits on a person. Revert and incident rates are the guardrail: “if they rise, we haven’t saved human time, we’ve only moved it.”

The layered mechanism is the most operational part of the post. Before a PR exists, work should be verifiable—tests, a build, or a screenshot against design—then reviewed by a fresh-context subagent. Boris Cherny is quoted calling a way for Claude to verify its work “probably the most important thing” for good results, estimating a two- to three-times better final outcome. Every PR gets a Claude review that reads the diff against the whole repo and the PR’s stated intent. In the research preview, the share of internal PRs getting substantive review comments rose from 16% to 54%. Findings are voted on by other agents before posting; style, diagnostics and cleanup noise were A/B-tested away on their own PRs.

Security review can block merges; so can lint, types, multi-platform tests and builds. Human approval scales by risk: low-risk changes (docs, copy, additive flagged work) need no human once automated and security checks are green; medium needs one person; high-risk and named sensitive paths need two. The change—via classifiers—sets the tier, not the author’s rank; only trusted authors get lighter paths; an agent’s approval never counts as a person’s. A merge queue re-runs checks against main; behavior usually ships behind flags to employees first, then latest, then stable. Incidents become a daily reviewer eval. Noise is scored by whether humans fixed, disputed, called intentional, or waved off a comment. One CLI PR ran more than fifty overnight rounds; git blame showed about 59% of what the reviewer kept flagging had been present since its first pass—hence a three-round cap, diff-only re-review thereafter, and no repeats of already-posted findings.

The strongest steelman for Anthropic is this: at their change volume, insisting on human line review raises both wait cost and miss cost; pushing routine defects into runnable verification and multi-agent voting, reserving people for direction, taste, billing/auto-updater/CI workflows and architecture that others will copy, and absorbing residual error with flags and rollback, is more honest than pretending every line was read. “You land it, you own it” and blast-radius thresholds try to keep automation from becoming ownerless shipping.

The hard gaps are equally clear. Every headline metric—90% Claude-written code, monorepo daily volume, PR doubling with >40% fewer reverts, 16%→54%, 59% preexisting repeats—is company-reported, without a public reproducible dashboard or external replication. The product still posts findings rather than approvals, yet their own repos already let owners merge some low-risk changes on automated checks alone, and “that line keeps moving”; they admit they do not yet have circuit breakers that tighten approval when revert or incident numbers move, and engineers want those before more code goes to zero-human merges. The falsifier is written into their own guardrail—but outside readers must currently trust post-hoc honesty rather than independently verify the live curves.

For engineering teams, the portable piece is not “hand review to Claude,” but an institutional checklist: CLAUDE.md / REVIEW.md, local /code-review and security review, risk tiers plus CODEOWNERS, merge queues and one-step rollback, incidents-as-eval, and an explicit ban on bots filling human approval slots. For buyers and compliance readers, the signal is that Anthropic has redefined “human in the loop” as a risk-configured accountability slot, not ritual line reading—if your contract or control still writes “a human reviewed the diff” as the compliance object, your “reviewed” and theirs are no longer the same noun.

On labor and organization, the shared-context section is sharp too: if machines do most of the reading, the team backfills understanding by asking Claude for slice explainers and watching open channels, not by reading trivial diffs. That may allocate attention better; it may also weaken apprenticeship paths that treated reading code as how you learn a system. The authors concede they have no formal substitute process.

Over the next six months, watch whether two curves still move together: falling human approvals and wait times, without rising reverts and incidents relative to the June–September improvement they claim. If zero-human merge paths widen while circuit breakers stay absent and incidents climb, “saved human time” is only a deferred bill. If the guardrail holds and the growing incident eval still suppresses noise, then the layered review Anthropic describes—at least inside its own repos—has rewritten review from a reading ritual into a risk-allocation system.

【No.2】