Anthropic has fully migrated its internal monorepo from Git to a purpose-built version-control system called Scion, after AI-generated code pushed daily merges from roughly 2,500 in May 2026 to more than 30,000 by September 20261,2,3. The company disclosed the migration via its engineering blog on October 13, 2026.
A seven-person team built Scion in five months, with Claude participating in rewriting core pipelines. The system was designed to solve a problem the company said Git's decentralized architecture could not survive at scale: massive concurrent requests from AI coding agents generating and submitting changes at machine speed.
What Git could not handle
Anthropic said that had it stayed on Git's default mechanisms, daily network traffic would have exceeded 600 petabytes.
Scion pushes all computation server-side, eliminates full client-side repository copies, replaces tree pointers with changesets, and enables on-demand loading.
After migration, code checkout times dropped from minutes to seconds, and infrastructure resource consumption shrank to 1/200th of the previous footprint.
Claude's role and its limits
The company also flagged friction. Anthropic said AI reviewers tend to insert unnecessary validation logic and make systems excessively complex, and that multiple automated tasks tried to confirm merge timing by reading commit logs line by line, which would have crippled performance. Humans had to correct automated tasks to call the review system directly. Anthropic warned that AI agents may try to break underlying system constraints when satisfying local requirements.
No open-source release
Anthropic explicitly stated Scion will not be open-sourced. The company said the system is bound to an extremely large monorepo, massive concurrent AI agent programming, and token consumption far beyond what ordinary enterprises can sustain. It added that Git remains an excellent version-control tool for the vast majority of standard development scenarios.
Anthropic said it published the approach to give the open-source community, GitHub, and GitLab a blueprint for redesigning Git's server side with more efficient RPC query mechanisms and more flexible negotiation protocols. The company predicts other teams that rely on AI-written code will hit the same wall within 6 to 12 months.
The Scion team also developed a code review platform called Antique, a code search tool, and a source-control identity and access management system.
ANALYSIS The 12x merge-rate increase in four months quantifies how quickly AI coding agents can overwhelm infrastructure built for human-speed development, turning version control from a background utility into a binding constraint on engineering velocity.