Anthropic Shifts Claude Code Review to Automated Checks Amid Volume Surge

Anthropic announced a fundamental shift in how the Claude Code team manages code review, moving the majority of reading tasks to machines to handle surging development volume. According to a blog post published on claude.dev, the team optimized its workflow to minimize human review time without increasing the rate of code reverts or production incidents.

This operational change addresses the rapid increase in code generation driven by AI agents. The Claude Code team reports that its own AI model writes approximately 90 percent of the codebase for the Claude Code product. Consequently, the volume of changes requires a review process that differs significantly from traditional manual, line-by-line inspection methods.

Automating the Review Process

The shift in strategy became necessary as development velocity accelerated across Anthropic’s infrastructure. In the company’s main monorepo, the number of code changes landing per day increased from about 2,500 in May to more than 30,000 in September of this year. While Claude Code resides in a separate repository, it faces similar pressures from high-velocity shipping.

From June to September, the number of merged pull requests for Claude Code more than doubled. During this same period, the share of those changes that had to be reverted fell by more than 40 percent. The team achieved these results by treating review as a continuous process rather than a final gate, similar to spellcheck operating while a document is being written.

Most changes are now reviewed and corrected by Claude at least once before a human engineer examines them. This automated first pass catches routine mistakes, allowing human reviewers to focus on higher-level questions and complex scrutiny. The team emphasizes that human time is a scarce resource, covering both the cost of reviewing and the cost of fixing broken code.

Context and Previous Reporting

This development follows earlier reports on Anthropic’s use of AI for infrastructure creation. As Threat Times reported on October 18, 2026, Anthropic’s AI model previously generated a clustered file system called MeshXFS by adapting XFS, raising questions about trusting AI-generated infrastructure for data storage. The current focus on code review processes complements those earlier findings by detailing how the company manages the quality and safety of AI-written code.

The team argues that asking whether a human read the code is an imprecise proxy for determining what could go wrong with a change. Instead, they evaluate which check—automated or human—is the right mechanism to catch specific issues. This approach preserves human attention for novel challenges that arise daily, ensuring that automated checks handle the bulk of routine verification.

What should you do?

Organizations adopting AI coding assistants should monitor their revert rates and incident frequencies to ensure that increased velocity does not compromise stability. Teams should implement automated review steps that run continuously, catching routine errors before human reviewers engage with the code. It is also important to maintain robust mechanisms for quickly undoing mistakes, as high-volume development increases the likelihood of errors reaching production. Finally, organizations should define clear boundaries for human review, focusing manual effort on complex logic and security implications rather than syntax or style.

What to do and how to stay safe: Anthropic

• Monitor revert rates and production incidents to ensure that increased code velocity does not degrade system stability.
• Implement automated checks that run continuously to catch routine errors before human reviewers examine the code.
• Establish clear protocols for quickly undoing mistakes, as high-volume development increases the risk of errors reaching production.
• Reserve human review time for complex logic and security implications rather than routine syntax or style checks.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How much of the Claude Code codebase is written by AI?

The Claude Code team states that Claude writes about 90 percent of the code for the Claude Code product.

How did code change volume change at Anthropic this year?

Daily code changes in Anthropic's monorepo rose from about 2,500 in May to more than 30,000 in September.

Did the team remove human code review entirely?

No, the team did not remove review but moved most reading to machines and optimized for human time and attention.