Anthropic built internal Git replacement Scion handling 30,000 daily merges

By Splitfeed EditorialLast updated Oct 15, 2026, 17:50 UTC

On October 13, 2026, Anthropic published a blog post describing how it built and migrated to an internal version control system called Scion, completing the switch away from Git in five months. The system now processes roughly 30,000 code merges per day, a volume Anthropic attributes to a tenfold surge in merge demand driven primarily by its AI coding agents running on Claude. Seven engineers are credited with shipping the replacement. Anthropic has no current plans to release Scion publicly, according to The New Stack. Hacker News threads have raised questions about how original the design actually is, pointing to a system called Mononoke as a close inspiration.

The numbers are what give this story weight beyond a routine infrastructure update. A tenfold increase in merge volume is not a marginal scaling problem; it suggests that AI coding agents are changing the fundamental tempo at which software development happens inside an organization. Git was designed around human contributors working at human pace, with occasional bursts. When autonomous agents generate and merge code continuously throughout the day, that assumption breaks down. Whether this pressure on version control is felt at the architecture level or mainly at the client layer is itself disputed, with some analysis arguing the strain lands on Git clients rather than Git's underlying design.

The Pro-AI camp is framing this as a milestone that illustrates AI productivity gains creating their own infrastructure demands. Accounts like stadolf and Jacob GPTProto emphasize the mismatch between tools built for human pace and the reality of agent-driven development, with stadolf summarizing it as "the bottleneck built its own cure." AlexTech.ai points to code management, not code writing, as the new constraint. PierreHenry.DEV is already asking when Scion might be released publicly for others to adopt. The Anti-AI camp has not yet published reactions to this story; typically, that camp would scrutinize whether the complexity and resource cost of building proprietary infrastructure around AI agents represents genuine progress or an avoidable consequence of over-reliance on those agents. The Middle Ground camp is more focused on the engineering lesson than on AI specifically: apiserpent argues the takeaway is simply to match tools to actual load rather than defaulting to industry standards, and Nick Gerakines notes that Git has long struggled with large-scale monorepo development regardless of AI involvement.

The most concrete open question is whether Anthropic will eventually release Scion. A public release would let other organizations evaluate whether the architecture genuinely solves agent-scale version control or whether the design debt flagged in Hacker News threads limits its broader applicability.