Anthropic describes how it checks Claude Code’s work


The company says Claude writes and reviews roughly 90% of the code produced by the team building its coding product, with automated checks supplying evidence for engineers.
Anthropic technical staff member Addy Osmani has described how the team building Claude Code checks work produced by multiple Claude agents. The system uses automated checks to give engineers evidence they can review without inspecting every change.
Anthropic says Claude writes and reviews roughly 90% of the code produced by the team building Claude Code.
The disclosure follows Anthropic’s earlier account of moving from human review at the end of development to continuous review, with Claude checking and fixing most changes at least once before a person is asked to look.
In answers to The New Stack, Anthropic said merged pull requests—proposals to add code changes—more than doubled from June to September 2026, while the share of changes reverted fell by more than 40%. The company did not say how much of that change came from verification or give separate figures for production defects and regressions.
Anthropic says it first runs repeatable checks such as builds, linters, which flag code issues, and tests of individual components and properties. These run as ordinary software once written. The company uses model review for judgments such as whether a change takes the right approach or misses a case no test covers.
Claude also exercises changes in a running application: it can use browser automation for web apps, connected simulators for mobile builds and local services for backend work. Engineers review evidence such as test results and screenshots. If a check still fails after three rounds, Claude hands the problem to an engineer.

A person is accountable for every change that ships, whoever typed it
To stop a proposed change from rewriting the rules used to judge it, the review system reads its REVIEW.md instructions from the base branch of the existing code rather than the proposed changes. A review applies only to the exact commit, or saved change, it examined.
The team’s checklist contains 39 reliability checks linked to past incidents or changes. Claude selects those relevant to a proposed change, typically running three to ten. Commit gates can require verification through /code-review or /verify, though developers can bypass a gate with an explanation; the team logs how each commit was checked, including when it was not.
Anthropic also bars Claude from skipping unreliable tests, inserting delays or retrying until a test passes. If it cannot identify the root cause of a failure, it deletes the test and documents why.
To grade its reviewers against past incidents and reversions, Anthropic recreates the repository as it stood at the time of review for each case. A reviewer gets credit only if its finding describes the actual bug; stable changes serve as controls, and a fixed test set is kept separate from examples used to tune the system. A real bug found but never flagged to an engineer counts as a miss.
Hacker News commenter OtherShrezzing argues that reusable Playwright tests could handle much of the verification without repeated model calls. Another commenter, xnorswap, warns that an agent might rewrite a failing test or add a workaround instead of fixing the underlying problem.
Anthropic has not given a specific figure for the added cost of running model-based verification or published the rates at which its reviewers raise false alarms or miss bugs. Without those figures, outside teams have limited data to compare their own implementations.
Anthropic says it wants most Claude Code changes to merge with at most one person involved, without increasing reversions or incidents. Repository owners already decide which low-risk changes can merge on automated checks alone; the company says its engineers want automatic controls that tighten approval if reversions or incidents rise before more code moves to merges without human involvement.
The unanimous November 19 vote replaces site-by-site licensing for an aviation band with a nationwide system. The final requirements for coverage at small-drone altitudes remain unconfirmed.
3 min

An analysis of nearly 10 million online ads found postings for more-exposed occupations fell about 10 percentage points further than those for less-exposed roles, but could not separate the technology’s effect from an economic slowdown.
3 min

The company says the radioactive materials license is the state’s first for a private fusion company. It permits fuel use at an Albuquerque research campus being built to test a system targeted for net energy gain by 2030.
2 min