On October 13, 2026, Anthropic published a long post on its official engineering blog detailing that the company had successfully migrated its large monolithic code repository (monorepo) from Git to its self-developed version control system Scion. The system was designed, implemented and fully migrated by a development team of only seven people in about five months, and was launched alongside complementary tools including the code review tool Antique, code search and an access control system. The direct driver of the migration was an explosive increase in code merges caused by AI agents—daily merge counts rose from around 2,500 in May 2026 to more than 30,000 in September.

Details

Anthropic noted that as models like Claude can autonomously iterate research cycles, micro-optimize complex programs and build services from scratch, coding itself is no longer the bottleneck; the real constraint becomes developer infrastructure designed for "human scale": source control, build systems and CI.

Merged changes to the Scion and Antique codebases per week, from January to September 2026.

The specific scale figures are striking:

The main branch history approaches 100 GB, and compressed working copies exceed 25 GB.
The CI system alone performs about 25 million checkouts per day; if full retrieval were done via traditional Git, daily traffic would exceed 600 PB.
The company previously relied on hundreds of internal Git mirrors, partial clones and snapshots as stopgaps, but with a 12-fold increase in merges these solutions became unsustainable.

Scion's core design principles contrast sharply with Git:

Always-online assumption: Git is distributed and supports offline full clones; Scion assumes clients (primarily AI agents and CI tasks) are always online and lets a server authority handle history queries, blame, merges, rebases and commit creation.
Fine-grained objects and incrementality: Changes are recorded as "changesets", storing only the modified files relative to parent nodes and content-addressed compressed blobs; clients retain a small local cache and fetch on demand.
Virtualized checkouts: Many tasks do not require a full checkout and can obtain diffs or specific files directly via API.
Scion draws on experiences from large-scale systems like Google Piper and Meta Sapling but is optimized for the high-frequency, fine-grained changes produced by AI agents.

The migration produced significant results:

Infrastructure CPU and memory usage dropped to about 1/200 of the previous Git mirror-based solution.
Checkout times for CI and research tasks fell from minutes to seconds (roughly two orders of magnitude improvement).
The CI system freed approximately 25% of CPU capacity.
Cache hit rates at all layers exceed 99%.

Notably, Claude itself played an important role in development. The team had Claude rewrite the Git-to-Scion data import pipeline; without extra instructions it produced a version with 20x lower memory usage and 4x faster performance.

Scion is currently for internal Anthropic use only. The company made clear it will not open source or offer it externally, emphasizing the system targets ultra-large scenarios such as "tens of thousands of engineers or agent clusters" and is unnecessary for ordinary projects.

Possible implications

1. Revealing new bottlenecks in AI-era developer infrastructure

When AI greatly reduces the cost and time of "writing code," downstream systems like version control, CI and code review can rapidly become the new constraints. Anthropic's case shows