Anthropic has fundamentally reshaped its developer infrastructure to handle a dramatic surge in code production driven by AI. The company reported that in May 2026, it merged approximately 2,500 code changes per day. By September, that number had grown to over 30,000 daily merges—a twelve-fold increase in four months.

According to Anthropic’s engineering blog, the rapid growth occurred because “coding is no longer the bottleneck to building systems or doing research.” Claude now autonomously iterates on research loops, optimizes complex programs, and builds new services at unprecedented speed. This success exposed a new constraint: developer infrastructure built for human-scale operations, particularly source control.
To address this challenge, Anthropic built Scion, a custom version control system designed from the ground up for their needs. The company fully migrated its monorepo from git in five months, concurrent with shipping other tools including a code review surface called Antique, code search capabilities, and a rich source control identity and access management system.
The scale of the problem was severe. Anthropic’s main branch history alone occupies nearly 100 GB, with a compressed working copy exceeding 25 GB. The company performs approximately 25 million code checkouts daily for CI, agent sandboxes, and research jobs—which would generate over 600 petabytes of git traffic daily if handled naively. Git’s distributed design, which assumes offline work and requires complete repository copies, became a liability rather than an asset.
Scion operates on fundamentally different principles. It assumes clients are always online—appropriate for agents and automated jobs—and implements strict client-server asymmetry. The server handles file history, blame, commits, merges, and rebases, while clients maintain only a bounded local cache. The system features a comprehensive API that allows tools and CI jobs to query source control without needing local repository copies.
Every component is independently horizontally scalable, with no fully materialized repository copy anywhere, even on servers. Commits are designed to be cheap, incremental, and lock-free, scaling with change size rather than total repository size. The working copy is optional and sparse by design, allowing CI jobs to request diffs or perform writes entirely through the API, eliminating unnecessary checkouts.
Anthropically acknowledges this workload is unusual but expects other organizations using AI coding tools will face similar pressures soon. The company notes that Google’s Piper and Meta’s Sapling provided valuable lessons for the Scion design.
Key facts
- Anthropic increased code merges from 2,500 daily (May 2026) to over 30,000 daily (September)—a 12-fold increase in four months
- The monorepo’s main branch history is nearly 100 GB; compressed working copy exceeds 25 GB
- Anthropic performs approximately 25 million code checkouts per day for CI alone
- Scion replaced git after a five-month migration of the entire monorepo
- Scion’s design prioritizes server-centric architecture, horizontal scalability, and sparse checkouts optimized for always-online clients
