Anthropic builds Scion monorepo as AI-generated code overwhelms Git workflows
Anthropic says its Claude models now produce so many code changes that its existing source-control and CI systems cannot keep pace. The company developed an internal centralized version-control system called Scion, described as a monorepo, after months of prompting and migration. It reported averaging about 2,500 merged changes per day in May 2026, rising to more than 30,000 per day by September.
Anthropic's internal system, Scion, emerged after months of prompting and a five-month migration. It is a centralized monorepo built to handle AI-agent workflows, with server-side queries exposed through APIs so tools can operate without local repositories. Claude also assisted in creating related tools, including a review interface, code search, and source-control identity management.
The scale is substantial: Anthropic's main branch is roughly 100 GB, and it performs about 25 million CI checkouts daily. A naive fetch could exceed 600 PB of traffic per day. Scion represents changes as changesets in a write-ahead log, storing changed files relative to a first parent and content-addressed pointers to compressed blobs.
The shift could affect software developers, DevOps teams, and code-review platforms most directly. As AI agents produce more changes, smaller organizations may struggle to match the custom infrastructure that large firms can build, potentially widening gaps in tooling and review capacity. Security and quality assurance teams may face harder tradeoffs between speed and scrutiny. If centralized, AI-oriented source control spreads, it could reshape how code is stored, reviewed, and audited across the industry.