Anthropic replaced Git with custom system Scion as AI-driven merges hit 30,000 a day

Anthropic outgrew Git

Anthropic has moved its codebase off Git onto an in-house source control system named Scion, according to an engineering write-up from the company, discussed by Simon Willison in a blog post that reached the Hacker News front page. The trigger was velocity: in May 2026 the company averaged roughly 2,500 merged code changes per day, and by September that figure had climbed past 30,000 — about a twelve-fold rise in four months.

The numbers behind the decision

Anthropic's stated requirements go well beyond what most organisations face. The company says the history of its main branch alone is nearly 100 GB, a compressed working copy tops 25 GB, and CI performs on the order of 25 million checkouts per day — before counting developer environments, agent sandboxes and research jobs. Served naively over Git, that would translate into more than 600 PB of traffic every day.

Centralized by design, with familiar ancestry

Willison reads Scion as a reinvention of centralized source control, closer in spirit to Subversion than to distributed systems like Git. Anthropic itself credits Google's Piper and Meta's Sapling as prior art, and Willison adds that the core "derived data" concept appears heavily indebted to Mononoke, the server component of Sapling. The write-up also introduces "Antique", Anthropic's new code review system.

The project was small and fast for its scope: according to Anthropic, seven people built Scion, and the migration off Git was completed in five months.

Offline support was deliberately left out. Asked about that omission on Hacker News, Scion engineer Nick G replied: "Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do."

What drove the surge

Anthropic's post includes a chart of merged changes over time. Willison had Claude edit that chart to overlay the public release dates of the company's 2026 models, and his interpretation is that the arrival of Fable/Mythos-class models is plausibly the biggest factor behind the speed-up. He cautions that the markers show public releases, not internal availability, so the correlation is suggestive rather than proven.

Adversarial review has limits

Alongside the infrastructure story, the write-up describes extensive use of adversarial review — models checking one another's work — which Anthropic found highly effective but sometimes myopic. The company says skilled engineers had to steer Claude away from over-engineering parts of the system in defence of hypothetical concerns. Willison notes this matches his own recent experiments pitting Claude and GPT against each other on personal projects.

Why it matters

Scion is a concrete signal that AI is reshaping developer infrastructure, not just the code developers write. Git was designed for human-paced contribution rates and offline, distributed workflows. When the heaviest committer is an always-online agent and daily merges run to five figures, those assumptions invert, and centralized systems in the lineage of Piper and Sapling start to look rational again — much as they did for Google and Meta at their own scale. If even a fraction of this change curve repeats across the industry, the next bottlenecks for AI-assisted engineering will be plumbing: source control, CI and review tooling built for agents rather than people.