Anthropic Built Its Own Git Replacement in 5 Months
Original: Redesigning source control at Anthropic
Why This Matters
Shows where AI-generated code volume breaks foundational dev infrastructure at scale.
Anthropic's daily code merges jumped from 2,500 in May 2026 to 30,000 by September 2026, forcing the company to abandon git and build Scion, an in-house version control system, in five months.
When AI-generated code volume grows 12x in four months, git breaks. That's what Anthropic discovered as Claude began autonomously running research loops and building services at machine scale. The company's monorepo history alone hit nearly 100 GB, with a compressed working copy exceeding 25 GB. CI alone required around 25 million checkouts per day—which, fetched naively, would have meant over 600 petabytes of daily git traffic.
To cope, Anthropic ran several hundred internal git mirrors at peak load. Seeds drifted further from the tip of main between refreshes, and partial clone operations like diff and blame grew increasingly fragile. The workarounds were running out of runway.
The solution was Scion, a purpose-built, centralized version control system designed from scratch for agent-scale workloads. Alongside it, the team shipped Antique (a code review surface), code search, a new IAM system for source control, and tooling for managing high volumes of reviewed code—all within the same five-month window. Anthropic notes a pointed irony: Claude's coding ability both caused the scaling crisis and powered the rapid development of the fix. The company acknowledges its workload is unusual today, but expects other heavy AI coding tool users to hit similar friction points soon.