Anthropic is turning its own internal code-review playbook into a product — and it’s aimed squarely at the messy reality of AI-era software development, where engineers ship more code than ever but have less and less time to read it carefully.

At the heart of the announcement is Code Review, a new feature inside Claude Code that throws a team of AI agents at every pull request instead of a single summariser bot. The system is modeled on the process Anthropic uses internally and is now rolling out in research preview for Claude Team and Enterprise customers, with reviews triggered automatically on GitHub PRs once an admin flips the switch. Anthropic is explicit about what this is and isn’t: it’s designed for depth, not speed, and it will not auto-approve your PRs — humans still own the green button.

The pitch starts from a problem that a lot of engineering orgs will recognise immediately. Inside Anthropic, code output per engineer has reportedly jumped 200 percent in the last year, thanks to coding assistants and agents. That productivity boost didn’t magically create more hours in the day for senior engineers to comb through diffs, so reviews became a bottleneck and many PRs got what the company bluntly calls “skims rather than deep reads.” Anthropic says that before Code Review, only 16 percent of PRs received substantive comments from human reviewers; after rolling the system out internally, that figure jumped to 54 percent. The company is betting that a lot of teams living with the same tension — more AI-generated code, thinner human attention — will be willing to pay for help.

Under the hood, Code Review behaves less like a single omniscient assistant and more like a panel of specialised reviewers. When a pull request opens, Claude Code dispatches multiple agents in parallel, each reading the diff and relevant context from a different angle, hunting for logic errors, edge cases, and fragile patterns that could ship subtle bugs. Those agents then cross‑check each other’s findings to filter out obvious false positives, and a final aggregator agent merges the results, deduplicates overlapping issues, and ranks them by severity before posting back to GitHub as one high‑signal summary comment plus a set of inline notes pinned to specific lines. Reviews scale with the change: large or complex PRs get more agents and a deeper pass, while tiny tweaks get a lighter touch, with the average review taking around 20 minutes in Anthropic’s testing.

Anthropic is already sharing a couple of “we would have shipped this bug” stories from its own usage and early customers. In one internal case, a seemingly routine one‑line change to a production service — the sort of diff that often gets rubber‑stamped — would have broken authentication entirely, but Code Review flagged it as critical before the PR was merged. In another, during a ZFS encryption refactor in TrueNAS’s open‑source middleware, the system surfaced a pre‑existing bug in adjacent code: a type mismatch that was silently wiping the encryption key cache on every sync, a problem that wasn’t actually introduced by the PR itself. Anthropic says that on big PRs changing more than 1,000 lines, 84 percent of reviews produce findings, with an average of 7.5 issues, whereas on small PRs under 50 lines, that drops to 31 percent and roughly half an issue on average — and less than 1 percent of those findings are marked as incorrect by engineers.

The company is also being upfront that this level of depth isn’t cheap. Code Review is billed on token usage and, in practice, translates into something like $15–25 per review on typical PRs, with costs scaling up alongside the size and complexity of the diff. That effectively makes it an opt‑in premium layer on top of the lighter, open‑source Claude Code GitHub Action, which Anthropic continues to offer for quick summaries and suggestions. To avoid surprise bills, admins get a few levers: organisation‑wide monthly spend caps, the ability to enable Code Review only on selected repos, and an analytics dashboard that tracks how many PRs were reviewed, what percentage of findings teams accepted, and the total cost.

From a developer’s point of view, the integration is deliberately boring — in a good way. Once a Team or Enterprise admin enables the feature in Claude Code settings, installs the GitHub app, and chooses the repositories to cover, reviews simply appear on new PRs with no extra configuration. The promise is that engineers keep using their normal GitHub workflow while Claude sits in the background as a very picky, very patient reviewer that never gets tired of re‑reading diff hunks. Crucially, Anthropic stresses that humans are still expected to make the final call on merges, and Code Review is not marketed as a replacement for human judgment but as a way to widen coverage when senior reviewers are stretched.

Zoomed out, Code Review is the latest in a string of Anthropic moves to position Claude Code as more than just a coding assistant that spits out snippets. The company has been talking up Claude as a reasoning‑first agent that can help with security analysis, legacy code modernization, and long‑context refactors, and this feature extends that story into the governance layer of software development. It also lands in the middle of an industry‑wide shift where “vibe coding” with AI tools is common, but formal review processes haven’t fully caught up with the volume of machine‑generated changes hitting production. For teams staring at ever‑growing PR queues and nervous about subtle regressions slipping through, Anthropic is essentially arguing that the only way to keep up with AI‑accelerated coding is to bring equally capable AI into the review room.