Scion: Anthropic’s lean Git alternative for agentic workflows
Anthropic introduces Scion, a server-based Git alternative. Always-online agents create enormous demand, pushing Git to its limits.
AI / iX
In a blog post, AI company Anthropic describes a homegrown version control system with a client-server architecture, developed after Git, the industry standard, reached its limits at the company. Anthropic does not plan to release the tool, however, because it is too closely tied to the company’s internal structures. The post is intended more as a case study.
Anthropic cites the growing need for code contributions (merges) driven by the work of agents as the reason for building its own system. Git, designed for offline development, could no longer scale sufficiently. The architecture of the new system, called Scion, is instead designed specifically for automated and agentic clients that are always online.
According to the blog, the Scion infrastructure requires 200 times less CPU power and memory than the previous Git mirror infrastructure. CI and research jobs check out code two orders of magnitude faster — in seconds rather than minutes (see Fig. 1).
Speed comparison between Git and Scion (Fig. 1).
(Image: Anthropic)
The company increased the number of merges from an average of around 2,500 per day in May to more than 30,000 per day in September (Fig. 2). “We don’t expect this rate of growth to slow down anytime soon,” the blog says.
(Image: Anthropic)
Alongside the core version control system, the AI company has built a review interface called Antique, as well as other supporting tools.
Git reaches its limits with agents
A Git system consists of multiple decentralized, independent clones. This structure made working with Anthropic’s monorepo increasingly difficult to manage. The main branch’s history was nearly 100 GB, and a compressed working copy exceeded 25 GB. On top of that, CI alone ran around 25 million checkouts per day, which would theoretically have meant more than 600 petabytes (about 600,000 terabytes) of daily data transfers with Git.
Specialized tools and workarounds reduced this load: Anthropic, for example, operated several hundred Git mirrors, which made functions such as diff and blame harder to coordinate (Fig. 3).
A typical Git checkout workflow using mirrors (Fig. 3).
(Image: Anthropic)
Scion’s basic design moves away from Git’s full-fledged clients to a client-server structure in which the connection is always online. Clients stay continuously synchronized and don’t need the entire repo as a base. CI systems also don’t need a copy of the repo; they retrieve the information they need through an API (Fig. 4).
(Image: Anthropic)
A key feature of Scion is its distributed server-side data storage. There is no materialized repository in one place; all data is distributed and scales horizontally. Unlike Git, Scion has no packfiles. A small number of servers is enough, with each making its functions available in all directions through an API.
Commits, rebases, and merges take place on the servers, so network traffic doesn’t face a large push load. Commits take the form of change sets, which record only changes to specific files rather than sending the complete modified files (Fig. 5).
(Image: Anthropic)
Videos by heise
Engineers remain valuable
A seven-person team developed Scion with Claude’s help in five months. One lesson Anthropic reports is that LLMs mean small teams need fewer people for migrations. Engineers are no less valuable as a result: “But our engineers still supplied the invaluable understanding of the high-level inter-system components and the design-level tradeoffs, making informed choices around the problems we wanted to solve and the scale we knew we needed to achieve.” They often had to correct the agents' work.
AI (adversarial LLMs) and people reviewed every change, focusing on the intent behind it rather than examining every line of code individually.
No open-source release
Anthropic does not plan to release Scion because, the company says, it is closely tied to its internal structures and therefore cannot be transferred elsewhere. Instead, the report aims to share the underlying lessons with other companies that have large development departments. They would then need to make the necessary adaptations themselves.
Other companies have also built their own version control systems, including Google with Piper and Meta with Sapling. In the latter case, even some core components are open source, though not the complete system.
(who)
