AI-led coding is also requiring a rewrite of some critical coding infrastructure.
Anthropic has built and rolled out its own version control system, moving its entire codebase off Git in just five months. The company says the switch became unavoidable once Claude’s own coding output overwhelmed the infrastructure that was built to handle human-paced development.
In a new engineering blog post, Anthropic revealed that in May 2026 it was merging around 2,500 code changes a day into its main repository. By September, that number had crossed 30,000 a day, more than a twelvefold jump in four months. The company says it doesn’t expect the growth to taper off either, since removing one bottleneck in the pipeline tends to immediately expose the next one, and the appetite for merging more code keeps climbing.
Git Wasn’t Built For Machines That Never Sleep
Git’s core design assumes every clone can work as a complete, offline copy of a repository, which is part of why it became the industry default. But Anthropic’s monorepo has grown so large that this design works against it: the history of its main branch alone is nearly 100 GB, and even a compressed working copy runs past 25 GB. The company was running roughly 25 million code checkouts a day for CI, agent sandboxes, and research jobs alone, which would have added up to more than 600 petabytes of Git traffic daily if fetched the standard way.
Anthropic had already pushed every conventional workaround to its limits, including running several hundred internal Git mirrors to spread the load. But as the merge rate kept climbing, these mirrors became harder to keep in sync, and operations like diff and blame grew increasingly fragile on partial clones. The underlying issue, the company says, is that its clients are no longer occasional human laptops working offline; they are agents and automated jobs that are online by default, which flips several of Git’s founding assumptions on their head.
Enter Scion
The replacement, called Scion, treats the server as the single source of truth instead of giving every client a full copy of the repository. Clients fetch only what they need, the server handles history, blame, merges, and rebases, and nothing on the system requires a fully materialized copy of the repo anywhere, not even centrally. Commits, rebases, and merges all happen server-side with no local pushing involved, and checkouts are sparse by default so a job only pulls the files it actually touches.
The payoff has been substantial. Anthropic says its Scion infrastructure footprint is now 200 times smaller than its old Git mirroring setup in terms of CPU and memory, CI and research jobs check out code in seconds instead of minutes, and CI alone freed up 25% of CPU capacity across the fleet.
Alongside Scion, the same small team also built a code review tool called Antique, a code search system, and a permissions layer for the monorepo, all within the same five-month window. Perhaps the most striking detail in the writeup is that a team of just seven engineers built all of this while the company’s daily merge volume kept climbing, a sign of how far Anthropic’s AI-driven software development practices have shifted the economics of building internal infrastructure.
Claude Built The Tool That Claude’s Own Code Growth Demanded
The irony Anthropic leans into is that Claude’s coding ability is both the cause of the problem and the reason the company could fix it so fast. The team notes that an early handwritten prototype of its Git-to-Scion migration pipeline was later handed to an internal snapshot of Claude Mythos Preview, which rewrote it with no further instructions into a version that used 20 times less memory and ran four times faster, while preserving every implicit invariant in the original ingestion logic.
Engineers on the project say they’d now rate Claude as a better code writer than themselves, but their own role didn’t disappear, it shifted. They described spending more time steering Claude away from shortcuts that would quietly violate system constraints, such as attempts to brute-force commit lookups by scanning the log line by line instead of querying the review tool properly, and reining in overly defensive code that resulted from Anthropic’s internal adversarial review culture.
A Narrow Fix, Not A Git Replacement For Everyone
Anthropic is careful to frame Scion as a solution to a problem that is, for now, mostly its own. The company says Scion only makes sense for organizations running truly massive monorepos, either with tens of thousands of engineers or swarms of agents generating code around the clock, and that it would be impractical for most companies to adopt wholesale since it requires its own bespoke code review, search, and CI stack to go with it.
Still, Anthropic expects this class of scaling problem to become more common over the next six to twelve months as more companies lean on agentic coding tools, and says it’s sharing the Scion writeup so that ideas from it, like server-side RPCs and smarter negotiation protocols, can be folded into Git itself rather than forcing every company to solve the same problem independently.