Anthropic reveals 90% of Claude Code is written by AI

The shift to AI authorship

Anthropic recently disclosed that about 90% of the code in its Claude Code tool was written by Claude itself. This transition marks a significant shift in authorship from human engineers to machine models. As the AI takes on more of the heavy lifting, the sheer volume of code changes has increased dramatically.

The scale of these changes is evident in the company's internal metrics. From June to September of this year, the number of pull requests merged into the codebase more than doubled. The amount of daily merged changes in the primary codebase jumped from approximately 2,500 in May to over 30,000 in September.

Despite this massive increase in volume, the team reported a counterintuitive success. The share of pull requests requiring a rollback fell by more than 40% relative to the previous period. This suggests that the automated systems are becoming more reliable as they scale.

A six-layer review system

Because manual human inspection of every line of code is impossible at this volume, Anthropic implemented a six-layer review process. The first gate happens before a pull request is even created, requiring Claude to run tests and capture screenshots. A sub-agent with fresh context then specifically reviews the code differences.

The second gate involves multiple AI agents hunting for bugs in parallel. These agents vote on issues, and only those meeting a specific threshold are posted as comments. This system successfully increased the share of substantive review comments from 16% to 54% since its launch in March.

The third and fourth gates focus on security and continuous integration. A pull request cannot be merged if any security findings remain open without a written justification. The continuous integration gate runs dozens of checks, such as linting and type checks, before a human ever sees the code. Recurring problems are documented in specific files to inform future reviews.

Risk-based human approval

The fifth gate introduces human approval based on a tiered risk system. Low-risk changes, such as documentation or copy, can skip human approval if they pass automated and security checks. These changes are often hidden behind feature flags to limit their impact during the initial rollout.

Ordinary product changes and sensitive areas like the auto-updater or system prompts require at least one human approver. The highest-risk code paths require two human sign-offs. This ensures that humans remain in the loop for the most critical parts of the infrastructure while letting the AI handle the routine work.

The final gate is the merge and release queue. The system re-verifies all checks against the most recent version of the main branch to ensure that individual changes do not break the system when combined. If something goes wrong, flipping a master switch turns the feature off without needing to change code or redeploy the system.

Lessons from automation

Automation is not perfect and can lead to strange behaviors. In one instance, a pull request for a command-line tool went through over 50 rounds of review and fixes overnight before a human intervened. The team found that 59% of the issues the reviewer flagged were actually in old code that existed before the review began, rather than new bugs.

Another bug appeared in a session cleanup optimization launched in March. It caused Claude to become repetitive and forgetful because it ran the cleanup on every turn after a conversation resumed. The issue persisted until April 10, and a later review showed that the Opus 4.7 model found the bug while Opus 4.6 did not, highlighting how model capability and context matter for detection.

To improve, the team updated rules to prompt authors to rethink their approach if a PR fails three rounds. They also limited reviews to only newly added changes and ensured the same class of issue is only raised once. These changes helped increase the share of internal PRs passing after a single round of review from 65% to 78%.

Expanding AI coordination

Anthropic is also applying these principles to drug discovery through its acquisition of Mithril Medicine in July 2026. Before the deal, the company had only one full-time employee who used Claude to coordinate contracts and data management with 11 external research institutes. This coordination helped achieve positive mouse-study results for a rare muscular dystrophy drug candidate in under 18 months.

The company disclosed that Claude made six errors during this research process. Some of these errors, including two data-verification mistakes, remained in emails sent to external institutes for more than four days despite human confirmation. This highlights the ongoing need for human oversight in complex coordination tasks.

The logic remains consistent across both software and medicine: drive down the cost of repetitive labor while keeping human judgment for accountability. For the software industry, this means engineering teams may adjust how they staff and assess performance as AI takes over more execution tasks.