Anthropic replaces git with in-house Scion as daily merges jump from 2,500 to 30,000+ in four months

Why it matters

【Key points】

Anthropic's engineering blog (Oct 13) says daily merges to its monorepo rose from about 2,500 in May 2026 to more than 30,000 in September, mostly Claude-written code. Git became the main bottleneck: about 100 GB of history, working copies over 25 GB, and about 25 million CI checkouts a day.

In five months Anthropic built Scion, its own version control system, and moved the monorepo off git. It assumes clients are always-online agents: commits and merges happen on the server, and checkouts are sparse or on demand. Anthropic reports 200x less CPU and memory than its old git mirrors, checkouts in seconds instead of minutes, and 25% less CI CPU. An early internal Claude Mythos Preview rewrote the ingestion pipeline to use 20x less memory and run 4x faster.

【Verifiable】

Post: https://www.anthropic.com/engineering/redesigning-source-control

@AnthropicAI: https://x.com/AnthropicAI/status/2110047081414987957 (about 218K views and 3,567 likes when fetched)

【Our read】

1. This is the first monthly measure a frontier lab has published of agent-written code: more than 10x in four months. The bottleneck has moved from writing code to infrastructure built for humans: source control, builds and CI. Other heavy users of AI coding will likely hit the same wall.

2. Much of this is self-reported. Merge counts are not quality, and engineers rating Claude a better coder is self-assessment. The more useful detail is how the role changed: engineers often had to tell Claude "no" and now review the intent of a change, not every line.

3. It bears on the debate over AI accelerating AI research. Watch whether Scion is open-sourced and whether other labs publish similar numbers.