Anthropic builds Scion, an in-house Git alternative for AI agents

Anthropic has detailed Scion, a version control system it built internally after Git hit its limits at the company. In a blog post, the AI developer said the surge in code contributions made by AI agents meant Git, designed for offline development, could no longer scale. Scion uses a client-server architecture aimed at automated and agentic clients that stay online at all times.

According to Anthropic, the new infrastructure uses 200 times less CPU and memory than the Git mirror setup it replaced. CI and research jobs now check out code about two orders of magnitude faster, in seconds rather than minutes. Daily merges rose from an average of around 2,500 in May to more than 30,000 in September, and the company wrote: "We don't expect this growth rate to slow down anytime soon."

Why Git broke down

Git relies on independent, decentralised clones, a model that grew increasingly unwieldy for Anthropic's monorepo. The history of the main branch approached 100 gigabytes, and a compressed working copy exceeded 25 gigabytes. CI alone performed roughly 25 million checkouts a day, which Anthropic said would theoretically have meant more than 600 petabytes of daily data transfer under Git. The company ran several hundred Git mirrors to ease the load, but that made features such as diff and blame harder to coordinate.

Under Scion, clients do not hold a full copy of the repository and remain continuously synchronised with the server. CI systems fetch what they need through an API. Server-side data is distributed and horizontally scalable rather than materialised in one place, and Git-style packfiles do not exist. Commits, rebases and merges run on the servers, with commits stored as change sets that record only the modifications to specific files. Anthropic also built a review interface called Antique and other supporting tools.

Built with Claude in five months

Anthropic said a seven-person team developed Scion in five months with help from Claude. The company concluded that migrations assisted by large language models need fewer people, but stressed that engineers remained essential for cross-system design trade-offs and often had to correct the agents' output. Every change was reviewed by adversarial LLMs and by humans, who checked the intent of a change rather than each line of code.

Anthropic will not release Scion as open source, saying it is too tightly bound to internal structures to be portable. The company framed the post as a way to share core lessons with other organisations running large engineering teams, which would need to build their own adaptations. Other tech giants have taken similar paths: Google uses its Piper system, while Meta built Sapling, parts of which are open source.