Anthropic Replaces Git With Its Own Source-Control System as AI Coding Surges

Anthropic has announced that it has replaced Git with a proprietary source-control system called Scion, following a dramatic increase in the volume of software changes generated by its engineers and AI coding agents. In an engineering report published on October 13, 2026, the company explained that its daily code merges increased from approximately 2,500 in May to more than 30,000 in September. The growth created a bottleneck in the infrastructure used to manage code changes, run tests and maintain its large software repository. Anthropic said it built Scion and migrated its main repository within five months. The company also developed supporting tools, including a code-review interface called Antique, as part of its effort to adapt its software-development process to increasingly autonomous AI systems.

The main problem was the growing size of Anthropic’s software repository and the demands placed on its existing Git infrastructure. According to the company, the history of its main development branch approaches 100 gigabytes, while a compressed working copy exceeds 25 gigabytes. Its continuous-integration systems perform approximately 25 million code checkouts daily, alongside additional activity from developer environments, AI-agent sandboxes and research jobs. Maintaining hundreds of internal Git mirrors helped manage the workload, but the increasing frequency of code changes made those mirrors more difficult to operate efficiently. Anthropic concluded that conventional approaches to source control were becoming a significant obstacle to its development process.

Scion uses a different architecture from Git. Git is designed around distributed repositories, allowing developers to maintain complete copies of projects and perform many operations locally, even without an internet connection. Scion instead assumes that most of its clients are AI agents and automated jobs that remain connected to central infrastructure. Its server handles operations such as code-history queries, commits, merges and rebasing, while clients retrieve only the information they need. The system also supports sparse and virtualised checkouts, reducing the amount of code that must be downloaded to perform a particular task. Anthropic says this approach allows its infrastructure to scale more efficiently as the number of automated coding operations increases.

Artificial intelligence also played a direct role in developing the replacement system. Anthropic said an early version of its Git-to-Scion data-import pipeline had initially been written largely by engineers. When the company tested the code with an internal version of Claude Mythos Preview, the AI model rewrote the pipeline and produced an implementation that used approximately 20 times less memory and ran four times faster, according to Anthropic. The company said the model preserved important requirements concerning the ordering of data ingestion and the underlying data structures. This example illustrates how AI is increasingly being used not only to write application code but also to redesign the infrastructure that supports software development. However, these performance figures come from Anthropic’s own report and have not been independently verified.

Anthropic’s experience could offer an early indication of challenges other companies may encounter as AI coding agents become more widely used. More automated coding can increase the volume of proposed changes, but organisations must still review those changes, run tests, resolve conflicts and maintain reliable release processes. If software-generation capacity grows faster than the infrastructure responsible for managing and validating code, development teams may face new bottlenecks. Anthropic believes some of these problems could become more common over the next six to 12 months. The company has therefore shared its engineering findings to encourage GitHub, GitLab and the open-source community to explore improvements such as more capable server-side operations and more efficient handling of large repositories.

Despite the announcement, Anthropic does not intend Scion to replace Git for general software development. The system was designed around the company’s unusually large repository and heavy use of AI agents, and adopting it elsewhere would require changes to related tools, including code review, search and continuous integration. Anthropic has not presented Scion as a broadly available replacement for existing version-control platforms. Instead, the company hopes other developers will learn from its experience and adapt selected ideas to their own systems. The development highlights a broader transformation in the technology industry: as AI systems generate more code, the tools used to test, organise, review and deploy that code may need to evolve alongside them.