Anthropic has replaced Git as the source-control system for its own monorepository with Scion, an internal platform built for a development workflow in which AI agents produce large numbers of changes. In an engineering account published on October 13 and updated October 15, the company says it completed the migration in five months. This is a report about Anthropic's internal infrastructure, not the launch of a tool that developers can download.
Why Anthropic changed its source control
Anthropic says its monorepo went from roughly 2,500 changes merged per day in May 2026 to more than 30,000 per day by September. It also reports about 25 million code checkouts per day for continuous integration, before counting developer workspaces and agent sandboxes. According to the company, the history of its main branch approaches 100 GB and a compressed working copy exceeds 25 GB.
Git can handle large repositories with shallow clones, sparse checkouts, mirrors and prebuilt snapshots. Anthropic says it had used those techniques, including hundreds of internal mirrors, but the operational cost kept rising with its merge rate. Its central observation was that most clients were online agents or automated jobs: they did not need the full local history and offline autonomy that make Git valuable for many other teams.
How Scion works
Scion makes the server responsible for operations such as creating commits, merging, rebasing, calculating diffs and answering history queries. A client keeps a bounded local cache and fetches only the files and metadata it needs. The design supports sparse or virtualized working copies, and some CI tasks can query the service without checking out a repository at all.
Its basic unit is a changeset: a server-created record of changed files and pointers to their contents in object storage. Separate derived data structures answer questions about ancestry, file history and line ownership. Anthropic says these components scale independently and avoid maintaining complete Git copies throughout its infrastructure.
The company reports that Scion's infrastructure now uses about 200 times less CPU and memory than its previous Git mirroring setup, that CI and research checkouts fell from minutes to seconds, and that CI gained 25% in CPU capacity. Those are Anthropic's measurements on its own workload; they do not establish the same gains for a typical repository.
What it means for Git users
Anthropic explicitly says Scion is not a suitable replacement for Git for general use. It depends on server availability and on a broader set of custom review, search and CI systems. The company expects few organizations to face its current scale immediately and says it is sharing design ideas that Git hosts might adapt, such as stronger server-side APIs and more efficient support for partial checkouts.
The case illustrates a change in developer infrastructure when coding agents sharply increase the volume of proposed code. It does not imply that teams should abandon Git: for most projects, Git's portable history and offline operation remain useful properties.