Anthropic published a detailed breakdown of Scion, its proprietary version control system that replaced Git for code created by Claude AI agents. The catalyst was an explosive growth in load: from May to September 2026, the average volume of daily merges increased from approximately 2,500 to over 30,000. The migration was executed by seven engineers over five months, and according to the company, Scion's infrastructure consumes 200 times less CPU and memory than the previous Git mirroring setup.

image

What Happened

The material was published in Anthropic's research blog under the title *Redesigning source control at Anthropic: The Scion System* and describes a system that already serves the company's monorepo in production. The scale that had to be sustained: the main branch history reached nearly 100 GB, the compressed working copy takes up more than 25 GB, and CI performs around 25 million checkouts per day—a naive fetch implementation would result in over 600 PB of Git traffic per day. Along with Scion, the code review tool Antique, code search, and IAM for the version control system were created. The company also reports accelerating CI checkout from minutes to seconds, saving 25% of CPU in CI, and achieving cache hit rates above 99%.

Context

Git was designed for human behavior: the distributed offline-first model considers a full clone of the history on the developer's machine the norm, while interaction with the server is reduced to manual fetch and push. When the primary author of changes becomes a machine working continuously, these assumptions break down: checkout and fetch are performed not episodically, but as a constant machine flow. Scion reverses this principle—"clients are always online." A commit is created on the server as a write-ahead log record (changeset), blame, diff, and ancestry are provided to clients via API, and data lives not in each working copy but in a hierarchy of caches: scion-daemon on the node, scion-cache in the region, and authoritative scion-servers on top of Spanner and cloud storage. In such a design, the working copy does not have to be self-sufficient: sparse and virtualized checkouts are used instead of a full clone.

Why This Matters for the Industry

This is the first public detailed breakdown of version control infrastructure designed for AI agents rather than humans. The main lesson for builders: the bottleneck in agent development has shifted from the model to the surrounding infrastructure—when code generation stopped limiting speed, version control, CI, and checkout became the limit. In the coming months, it is reasonable to expect VCS vendors, CI, and dev-tools startups to announce agent-scale features: server-side operations, API-first reviews, caching layers for CI, while companies with agent-heavy codebases will launch pilots of partial migrations or acceleration layers on top of Git and begin to rethink their Git stack entirely. Git, however, will likely remain the standard for small and medium-sized teams. Within a couple of years, a split in the toolchain into human-first and agent-first paths is likely, with the emergence of a class of "VCS for agents" products. All of this consists of expectations and hypotheses based on the trajectory of one company, not an established trend.

Why This Matters for Users

Scion cannot be used directly: in the materials, it is described as an internal Anthropic system with no mention of a product, open API, or open source. The value for the reader here is analytical. The article breaks down solutions that are rarely described publicly: blame as derived data, comparison with Git packfiles, the structure of the server change log, and the cache hierarchy. The migration case of critical infrastructure during active development is of particular interest—an example that such a replacement is feasible if the team has full control over the environment. The material is also useful for those following the structure of internal systems working with Claude. The practical first step is to read the original article and compare the metrics of your own agent pipeline: the number of merges and checkouts per day, the volume of fetch traffic—with the scales provided in it.

What Is Still Unknown / Limitations

All key results—CPU and memory savings compared to the previous Git mirroring, checkout acceleration, reduced load on CI, and cache hit rate—are claimed by Anthropic itself; there is no independent verification of these figures in open sources. Scion is described as an internal system: the published data does not mention productization, an open API, or open source, so it is currently impossible to use it outside of Anthropic. Forecasts about the appearance of agent-scale features among vendors and the split of the toolchain into human-first and agent-first paths are interpretations based on the trajectory of one company, not established facts.

Sources

Author

Look at AI, editorial team