Anthropic Replaces Git with Custom Scion System Following AI-Driven Code Surge
Anthropic has replaced Git in its main shared codebase with a system it built itself called Scion. This change came after the number of code changes made each day grew from about 2,500 in May to more than 30,000 by September. The rapid growth of AI coding agents that work on their own, powered by Claude, sped up software development so much that managing, reviewing, and testing changes became a bigger challenge than writing code. Over five months, a team of seven engineers built Scion and a related code review system called Antique. They used Claude to write large parts of the code, while human engineers guided and supervised the work.
Moving the Main Challenge from Writing Code to Building the Systems That Support It
The change happened because the work needed to support each software submission began to put too much pressure on Anthropic's internal developer systems. The company's automated testing systems retrieve code about 25 million times every day to check new changes. Keeping hundreds of local copies of the Git repository had become harder to manage and required more resources to keep those copies in sync as changes arrived so quickly.
Scion reduces this extra work by moving processing to central servers. This lets automated systems and engineers request only the specific files or information they need. Automated tools can check past changes or create comparisons between versions without downloading a complete local copy of the repository. After Scion was introduced, Anthropic said code retrieval times fell from minutes to seconds. The new system also reduced the CPU and memory resources needed to one two-hundredth of what the old Git mirroring system used.
Human Oversight and Design Rules
Although Claude wrote a large amount of the code, human engineers decided how the system should be designed, how its parts should work together, and how to handle unusual cases. Anthropic Chief Technology Officer Rahul Patil said on LinkedIn that the team had to make important design decisions. They often had to reject Claude's suggestions when they broke important design rules. In one example, Claude generated code that tried to search through an entire history of code changes to find information that was already available directly through the review platform.
Engineers have since shifted their attention away from checking every line of code one by one. Instead, they focus more on understanding the purpose of changes, security limits, and whether the overall design makes sense. To keep the system stable while code changes arrive so quickly, the team built separate testing environments, automated fuzz testing to find unexpected problems, and safeguards that limit how many requests the system can handle. Scion also provides detailed permissions for individual directories and review tools that can tell the difference between minor changes, such as edits to comments, and changes to the software's main business logic.
A System Built for a Specific Scale, Not a Universal Replacement
Anthropic stressed that Scion was designed specifically for its own demanding environment, which includes huge amounts of work by AI agents and an unusually large shared code repository. Using Scion at other companies would require major changes to their existing search systems, code review tools, and automated testing systems.
The company explained that Git remains the best version control system for typical software teams. Scion is not meant to replace Git for everyone, and it is not designed as a ready-to-use public developer product. Instead of releasing Scion as open-source software, Anthropic is sharing lessons from its design and the methods used to build it. Other teams can use these ideas with standard Git setups to handle more work centrally and reduce unnecessary data transfers.