At some point in the last few years, the conventional wisdom about distributed version control stopped fitting how AI development actually works. Git was designed around the assumption that developers are humans who are sometimes offline, and that every clone should be a complete, self-contained repository. Those assumptions made sense in 2005. They look different when most of your commits come from agents running in CI.
Anthropic’s engineering blog published a post today describing how they outgrew git and what they built to replace it. The system is called Scion. The timeline they describe is striking: their monorepo processed about 2,500 merges per day in May 2026. By September, that number had grown to over 30,000. Most of that traffic comes from AI agents and automated pipelines, not humans typing at terminals.
The first response was scaling out git infrastructure — running several hundred git mirrors to handle the load. That bought time, but the fundamental issue was that git’s architecture doesn’t separate “client needs a checkout” from “server stores the canonical state.” Every mirror had to hold a full copy of everything. As the repo grew and merge velocity increased, keeping hundreds of complete mirrors consistent became genuinely hard.
Scion takes the opposite design stance: clients are always online, so there’s no reason to build for offline use. The server handles history, blame, commits, and merges. Clients keep a small local cache and fetch on demand. Commits happen on the server and scale with the size of the change, not with the size of the repository. CI jobs and research pipelines that only need to query file contents can skip a local checkout entirely and just call the server.
The performance numbers they report are large. Scion’s infrastructure uses roughly 200 times less CPU and memory than the old git mirroring setup did. CI and research checkouts that previously took minutes now take seconds. CI alone recovered about 25% of its CPU capacity. Anthropic also moved away from a global merge queue — with this many automated commits, a single merge queue is a bottleneck by design — toward post-merge and pre-deploy testing.
Seven engineers built Scion and a companion code review tool they call Antique in about five months. They describe Claude as a significant contributor to the implementation, and make an observation worth noting: Claude often needed to be told “no.” Engineers’ roles shifted toward setting constraints and reviewing intent rather than writing most of the code. That’s become a recurring theme in AI-assisted large-scale engineering: the bottleneck moves from implementation to judgment about what to build and what to stop.
There’s a broader pattern here worth naming. AI-driven development is eating its own tail in an interesting way. The very tools being built to make agents more capable — faster CI, more reliable code indexing, better merge pipelines — are themselves being strained by agent activity. Anthropic’s monorepo didn’t grow because humans started committing faster; it grew because agents did. Scion is infrastructure built to support agents, built partly by agents, under load generated by agents. That loop is still accelerating.
Anthropic is explicit that Scion isn’t intended as a general git replacement. The design tradeoffs — centralized, always-online, no full local clone — don’t make sense for most projects. But they say they plan to share the ideas, and it seems worth watching whether git hosting platforms start absorbing some of them. The case for “server does the merge, client caches on demand” is stronger when the commit rate makes a distributed model impractical, and that condition is going to describe more and more repositories over the next few years.