Anthropic’s Claude Code team says Claude writes about 90 percent of Claude Code’s own code, and that from June to September merged pull requests to its codebase more than doubled while the share that had to be reverted fell by more than 40 percent, according to a post on the team’s blog published on 15 October 2026.
The post describes how the team changed code review to get there. It says the team “didn’t remove review”, but moved most of the reading to machines and invested in making mistakes fast and easy to undo.
Business Pill · THE AI AUDIT PROBLEM
A Business Engineer explainer on why AI work is hard to audit. It explains the concept in general and says nothing about Anthropic’s review process.
The key insight: As we read it, the Claude Code team is describing a change in what review is for: the post says machines now do most of the reading before and on every pull request, and people are kept for the decisions it lists, from risky code to whether a change should be built at all.
The post says most review now happens before a pull request exists. Claude is given something it can run, such as tests, a build or a screenshot to compare against a design, and a subagent then reviews the diff with fresh context.
Every pull request then gets a Claude review before a person looks. According to the post, several agents look for bugs at the same time, a separate group votes on each finding, and a finding is posted only if enough of them agree it is real. A separate automated security review can block a merge while any finding is still open.
The team says that when it launched the research preview of its code review product, the share of internal pull requests getting substantive review comments went from 16 percent to 54 percent. The post says Code Review for Claude Code is now generally available on Team and Enterprise plans.
The share of Anthropic’s internal pull requests getting substantive review comments went from 16 percent to 54 percent when the Claude Code team launched the research preview of its code review product, according to the team’s blog post of 15 October 2026. Anthropic’s own figures.
What Was Posted on X
Claude writes about 90% of Claude Code’s code.
From June to September, merged PRs more than doubled while the share reverted fell by more than 40%.
We wrote about how we changed code review to get there, and how to adopt it:
The Claude Code team’s own account, @ClaudeDevs, posted on 16 October that Claude writes about 90% of Claude Code’s code and that from June to September merged PRs more than doubled while the share reverted fell by more than 40%, linking the blog post this piece is based on.
Worth reading: How Anthropic reviews AI-generated code.
Claude writes about 90% of Claude Code’s code. From June to September, merged PRs more than doubled while the share reverted dropped by over 40%.
Here’s how they handle code review:
1. Before a PR: run tests and builds, then have a fresh agent review the diff.
2. Every PR: multiple agents hunt for bugs, others verify the findings, plus security review and CI.
3. Human approval depends on risk: 0 for eligible low-risk changes, 1 for medium, 2 for high.
4. New features roll out gradually behind feature flags, with fast rollback if something breaks.
5. Past incidents become daily evals to improve the AI reviewer.
The most useful insight is how they measure human cost.
It includes both time spent reviewing code and time spent fixing what bad code breaks.
Cutting review time doesn’t help if you end up spending more time fixing production incidents.
Developer Kai (@hqmank) summarised the post on 17 October in five steps, from review before a pull request to risk-based human approval and incident evals, and singled out how the team counts human cost as both reviewing time and time spent fixing what bad code breaks.
How Many People Approve a Change
Human approval is set by risk, the post says. Low-risk changes such as docs, copy and additive changes behind a flag need no human approval once automated review and security checks are green; ordinary product changes and sensitive areas such as caching, the auto-updater and system prompts need one person; the highest-risk code needs two.
“The change sets the tier, not the author,” the post says: a classifier reads the change and picks its risk tier. It adds that only authors on a repository’s trusted list get the lighter paths, and that an agent’s approval never counts as a person’s.
The goal, the team writes, is that most pull requests merge with at most one human in the loop, who can be the author, with no increase in reverts or incidents. It says it tracks human approvals per pull request, human comments and waiting time, with revert and incident rates as the guardrail.
What Still Breaks
The post says some mistakes still get through, so changes ship behind flags to Anthropic employees first and can be turned off without a deploy. It says the team does not yet have circuit breakers that tighten approval when revert and incident numbers move, and that engineers want them before more code moves to zero-human merges.
It also describes a cost of agents reviewing agents: one pull request went through more than fifty rounds of review and fixes overnight before its author stopped it. When the team ran git blame on what the reviewer kept flagging, it says about 59 percent was code that had been there since the reviewer’s first pass. The team now caps review at three rounds before asking the author to reconsider the approach.
As we read it, the number the team watches is human time rather than lines read: the post says it tracks approvals per pull request, comments and waiting time, and treats revert and incident rates as the guardrail.
As we read it, the approval rule is tied to the change, not the person: the post says a classifier sets the risk tier, low-risk changes can merge with no human approval once checks pass, and an agent’s approval never counts as a person’s.
As we read it, the team also names the limits: the post says it has no circuit breakers yet that tighten approval when revert and incident numbers move, and that one pull request ran more than fifty review rounds overnight before a three-round cap was added.
Claude Code team blog, 15 October 2026
“We didn’t remove review. But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes.”
Three Implications
ENGINEERING LEADERS The post says the team tracks human approvals per pull request, human comments and how long a pull request waits on a person, with revert and incident rates as the guardrail.
TEAMS ADOPTING AI REVIEW The post lists steps such as a short CLAUDE.md, a REVIEW.md written by code owners, an AI reviewer on one repository first, and a cap on review rounds.
SECURITY AND COMPLIANCE OWNERS The post says a pull request cannot merge while any bug or security finding is still open, and that some paths, such as CI workflow files, always need a person.
The framework argues that the next jump “is not a better prompt. It is an operating system that wraps your AI”, a harness that turns the model’s raw capability into something a team can repeat.
As we read it, the Claude Code team’s review layers, CLAUDE.md and REVIEW.md files, risk tiers and incident evals are that kind of wrapper, built around the model that writes most of its code. The framework was published in April 2026 and does not discuss this post.
What Is Not Established
The post does not give the absolute number of merged pull requests or reverts, only the change from June to September. It describes the Claude Code team’s own repositories, not every team at Anthropic or customers’ codebases.
We did not test Anthropic’s review process or its code review product, and we did not contact Anthropic.
Business Engineer Framework
AI & The Harness Theory
A Business Engineer analysis of the operating system that wraps an AI model so a team can rely on it.
Anthropic’s Claude Code team says Claude writes about 90 percent of Claude Code’s code, and that merged pull requests more than doubled from June to September while the revert share fell by more than 40 percent.
It credits a review process that runs AI review before and on every pull request, sets human approvals by risk at zero, one or two people, and relies on flags and fast rollback for what still gets through.
A note on sourcing. This piece is based on the Claude Code team’s blog post on code review, published on 15 October 2026, and the @ClaudeDevs post that links it, both read on 17 October 2026. All figures are Anthropic’s own. We did not test the process or the product, and we did not contact Anthropic. The references to our earlier reporting link to our own pieces.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
