Redesigning Source Control at Anthropic (www.anthropic.com)

Anthropic has unveiled a ground-breaking in-house version control system called Scion, designed to address the challenges brought by the rapid growth of code changes in their AI-driven development environment. With Anthropic now merging over 30,000 code changes daily—a significant leap from an average of 2,500 in May 2026—traditional systems like Git have proven inadequate, particularly for managing their enormous monorepo, which has created bottlenecks in source control and continuous integration (CI). Scion was built from the ground up to optimize for an environment where clients (AI agents) are always online, making it fundamentally different from the typical decentralized approach of Git. Key technical innovations in Scion include a client-server setup that minimizes local data requirements and encourages efficient on-demand data fetching directly from a centralized server. This design allows for faster commits and more scalable operations by recording changesets without the need for traditional Git refs or packfiles, thus eliminating the complexity associated with large-scale data management. Scion's comprehensive API exposes source control queries efficiently, fostering the development of novel applications and enhancing developer productivity. As AI-enhanced coding becomes more prevalent, Anthropic expects other organizations to experience similar scaling issues, making their insights on source control architecture increasingly relevant to the wider AI/ML community.