Anthropic reinvents source control as Claude floods the world with code

Anthropic's Claude models have generated such an overwhelming amount of code that the company's coding infrastructure struggled to keep pace. This issue stems from the AI coding imperative and the availability of all-you-can-eat inference. Microsoft's GitHub noted growing availability problems due to the exponential increase in pull requests and commits.

Researchers have also observed the code flood. For Anthropic, writing code is no longer a production bottleneck; the real challenge is managing the code after it has been written, through source control, build systems, and continuous integration.

As of September, Anthropic was merging over 30,000 code changes daily, a significant increase from around 2,500 in May 2026. Anthropic does not expect this growth rate to slow down, as the demand for merging code is practically infinite, and removing bottlenecks often leads to even more code production. After months of model prompting and migration, Anthropic developed a homemade version control system called Scion, a monorepo scheme resembling Subversion.

While Claude's code-writing capabilities created the problem, they also provided a solution through the creation of scalable code infrastructure tools, including Scion, Antique (a code review interface), a code search tool, and a source control IAM system.

Anthropic emphasizes that the decentralized approach of Git is suitable for most code repositories. However, their main branch is approximately 100 GB, with a compressed working copy exceeding a quarter of that. The company performs around 25 million code checkouts daily for CI, not accounting for developer environments, agent sandboxes, and research jobs.

Fetching data naively would result in over 600 PB of git traffic per day. To address this, Anthropic designed Scion to cater to AI agents, with most work handled by the server, minimizing client-side operations and enhancing server speed. Scion's changeset, a write-ahead log entry, stores most of what a git commit would, but lists the changed files relative to its first parent and includes content-addressed pointers to compressed file contents in blob storage.

Development on Scion began in January, and when Claude Mythos Preview became accessible internally, Anthropic's code experts instructed the model to rewrite an early Git-to-Scion ingestion pipeline. The result was a version using 20 times less memory and running four times faster. With ample venture funding, such optimizations are justified.

Anthropic's Claude engineers acknowledge the model's superior code-writing abilities but maintain a higher-level understanding, often guiding Claude to avoid violating core system constraints. While Anthropic's bespoke system is not open source, the company believes its code scaling challenges will become widespread, and they wish to share their approach with other organizations.