An online-first version control system in this agentic era
Anthropic recently shared an article explaining a version control system (VCS) that they have been using internally to overcome their own limitations around managing their humongous monorepo.
The VCS is called Scion and it flips the way how a traditional VCS like Git work. For instance, Git is built around a decentralized model where every client act as a full-fledged, offline-capable server with a complete copy of the repository. This way all these clients can work offline on their own respective copies.
On the contrary, Scion assumes that all its primary clients (AI agents and automated jobs) are always online. The server that hosts the monorepo handles history, blame, commits, merges, and rebases. And so, the client doesn’t need to compute complex graph relationships locally. Instead, it maintains a bounded local object cache and relies on the server being reachable.
But why did they invent this new tool in the first place?
The answer lies into their workflows that involved exponential merge volume. In May 2026, Anthropic averaged around 2,500 code changes merged per day. By September, that number skyrocketed to over 30,000 merges per day. Traditional git workflows, branch management, and human-scale merge queues simply could not keep up with that velocity.
In addition to this, Anthropic uses a massive monorepo. The history of the main branch alone grew to nearly 100 GB, with a compressed working copy exceeding 25 GB. In Git, every client clones a complete repository and computes history, diffs, and blame locally. Scaling this up for thousands of continuous integration (CI) jobs, agent sandboxes, and research tasks meant dealing with tens of millions of checkouts a day—amounting to an unmanageable 600 PB of git traffic daily.
Scion excels in mitigating this limitation since clients are always connected, CI jobs and research tasks don’t need to fetch or maintain multi-gigabyte local repositories. They query the server directly via APIs or fetch sparse/virtualized files on demand.
So, a job, for instance, will only ask for the file it needs when it needs them. So, a job gets its code in seconds instead of minutes.
Anthropic says, Scion uses 200 times less computing power than the old git mirrors and it freed up a quarter of all the computers they use for testing which is a huge win considering the amount of merges they do each day.
Even though all these sounds too good to be true, Anthropic is not releasing Scion publicly since packaging it into a general-purpose, public product would require massive engineering overhead to make it stable, configurable, and usable outside of their specific infrastructure.
They are sharing the ideas around it since they expect many more teams to hit the same limits soon in this new era where extreme agentic workloads handling tens of thousands of automated merges per day in a massive monorepo simply becomes inevitable.
👋 Hi there! This is Amit, again. I write articles about all things web development. If you enjoy my work (the articles, the open-source projects, my general demeanour... anything really), consider leaving a tip & supporting the site. Your support is incredibly appreciated!
