Redesigning source control at Anthropic
As our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.
This article details Anthropic's development of Scion, an in-house version control system designed to address scalability challenges caused by their AI-driven code development workflows.
What happened
Anthropic developed Scion, a custom version control system, to replace Git as their monorepo grew beyond 30,000 daily code merges. The system addresses limitations of traditional distributed version control by optimizing for always-online agents and large-scale code management.
Why it matters
For AI developers and infrastructure teams, this demonstrates how AI-assisted development creates new demands on tooling. The architectural decisions in Scion provide a blueprint for building systems that scale with AI-driven workflows.
- AI-driven workflows require rethinking traditional software infrastructure designs
- Server-centric architectures can outperform distributed models at scale
- Derived data structures enable efficient query capabilities without full repository copies
- AI tools can accelerate development of complex systems when properly integrated
The bigger picture
The system reduced infrastructure footprint by 200x, improved CI checkout speeds by 100x, and eliminated repository maintenance requirements, demonstrating practical benefits for AI-driven development pipelines.
Anthropic's Scion project demonstrates how AI-assisted development creates new infrastructure requirements, with lessons applicable to other organizations facing similar scaling challenges.
Technical details
- Changesets: Server-side units storing only modified files with content-addressed blobs
- Derived data structures: File blame tracking, path history, and commit graph indices computed from changesets
- Horizontal scalability: No single point of failure, with tiered caching architecture
- Sparse checkouts: Virtualized file access reducing disk I/O requirements
FAQ
What problem did Scion solve for Anthropic?
Scion addressed the limitations of Git as Anthropic's code merge rate increased from 2,500 to 30,000 daily merges, requiring a system optimized for agent-scale workloads.
How does Scion differ from Git?
Scion uses server-centric design, changeset-based data structures, and horizontal scaling to handle large monorepos more efficiently than Git's distributed model.
What are the key benefits of Scion?
Scion reduces infrastructure requirements by 200x, enables 100x faster CI checkouts, and eliminates repository maintenance through its scalable architecture.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Lobsters on Oct 13, 2026.
Comments
Sign in to join the discussion.