Anthropic deploys new source control system Scion in place of Git

Summary

Anthropic replaced Git with Scion, a proprietary, centralized source code management system, after AI agents increased the daily number of merged changes from approximately 2,500 to over 30,000 in four months, generating up to 600 PB of raw Git traffic per day. Developed by a seven-person team over five months, the system intentionally lacks offline support because its most intensive user, Claude, is continuously online.

Why it matters

This is a concrete, production-scale signal that AI-agent-driven development is challenging two decades of assumptions underlying distributed version control. The decision to replace Git with a centralized, server-centric architecture points to a structure that Simon Willison compared to Subversion. It demonstrates that when the primary writers and consumers of code are machines rather than humans, the tooling stack must be redesigned around constant connectivity and massive file download speeds.

Highlights

• The daily merge count increased approximately 12-fold between May and September 2026 (from 2,500 to over 30,000). This rate made Git's distributed model, which was struggling with 25 million file downloads per day, unmanageable.
• No offline mode: The decision is justified by Claude being the dominant user and inherently online. This represents a philosophical break from Git's human-centric, distributed design.
• While adversarial AI code review is 'incredibly effective,' it can be short-sighted. This pushes engineers to steer models away from over-engineering for hypothetical edge cases.

Critical take

The lack of offline support is a revealing rationale. It normalizes a world where the primary developer is a cloud-connected large language model (LLM). This approach may be suitable for Anthropic's internal workflows; however, it raises doubts about whether Scion is a general-purpose version control system or merely a pipeline specialized for the processing speed of AI agents. Additionally, a seven-person team delivering a system that replaces a 20-year-old, field-tested tool in just five months warrants a cautious perspective regarding long-term sustainability and whether the centralized model can withstand the growth of the codebase and agent fleet.