Anthropic migrates to Scion, a new source control system designed for AI agent scale

Anthropic migrated its internal monorepo from git to Scion, a custom source control system designed to handle the extreme volume of code generated by AI agents. The shift was driven by a surge in merge frequency, which jumped from approximately 2,500 daily merges in May 2026 to over 30,000 in September 2026, overwhelming traditional git infrastructure that struggled with a 100GB main branch history and massive CI checkout traffic.

Scion Architecture and Performance

Scion adopts a server-centric design where clients are always online and rely on the server for history, blame, and merge operations, eliminating the need for local repository maintenance or packfiles. Key architectural features include:

• Changesets: Units of work stored as write-ahead log entries with content-addressed pointers, avoiding the overhead of full tree pointers.
• Tiered Caching: A hierarchy involving scion-daemon, scion-cache, and scion-server ensures high availability and low latency, with cache hit rates exceeding 99%.
• Horizontal Scalability: Data is distributed without fully materialized copies, allowing independent component scaling.

The migration resulted in a 200x reduction in CPU and memory infrastructure compared to git mirroring, and CI checkout speeds improved by two orders of magnitude, dropping from minutes to seconds.

Development with Claude

A seven-person team built Scion and related tools like Antique (code review) in five months, heavily utilizing Claude Mythos Preview for development. Claude assisted in rewriting ingestion pipelines, reducing memory usage by 20x and increasing speed by 4x. The team also leveraged Claude for algorithmic optimizations, such as discovering missing indices for ancestry queries, a task that would typically take experts days or weeks.

Limitations and Future Outlook

Scion is not intended as a general replacement for git but is tailored for "truly massive monorepos" where agent swarms generate code at scale. Anthropic notes that Scion’s bespoke nature makes it difficult for other organizations to adopt directly. However, the company plans to share insights with the broader community, suggesting that git servers could adopt similar server-side RPCs and scaling strategies to better support AI-driven development workflows in the next 6–12 months.