Picture a toll booth built for morning commuters. One lane, one attendant, works fine for decades. Then a freight depot opens next door and runs around the clock, and suddenly the problem isn’t that the booth is badly built. The problem is that it was designed for a traffic pattern that no longer exists. That’s roughly what happened to Git inside Anthropic, and in October 2026 the company published how it responded: a custom version control system called Scion, which it has already migrated to internally.
I review tools for a living, and I’m usually the guy telling people not to rebuild their stack. So let me say upfront that this one has a number attached to it that’s hard to argue with.
The number that justifies the rewrite
In May 2026, Anthropic averaged around 2,500 code changes merged per day in its monorepo. By September, that figure was over 30,000 per day. Same company, roughly four months, more than a tenfold jump in merge volume.
That is the entire story in one statistic. Anthropic’s framing is that writing code stopped being the bottleneck. The bottleneck moved downstream into source control, build systems, and CI, which is exactly where you’d expect it to land once code generation gets cheap and merging doesn’t.
Git was designed around a human cadence. A developer pulls, works for a while, commits, pushes, goes to lunch. Agents don’t go to lunch. They’re always online, they operate in parallel, and they produce changes in a steady stream rather than in bursts around standups. When your traffic profile changes that much, tuning the existing system eventually stops paying off.
What Scion actually does differently
Based on what Anthropic has described, Scion is built for always-online agents and automated jobs rather than for people at laptops. The design choices it highlights:
- Incremental changesets instead of treating every change as a full-weight commit object in the Git sense.
- Scalable object caches, which matter enormously when thousands of jobs want overlapping slices of the same repository at the same time.
- Sparse or virtualized checkouts, so a job pulls the slice of the tree it needs rather than materializing the whole monorepo.
Anthropic reports substantially faster checkouts and a smaller infrastructure footprint than Git. The company also deployed cloud Tag instances and Slack channels to monitor the health of specific components, which tells you something about the operational reality here: this isn’t a tool you install and forget. It’s infrastructure with a pager attached.
The honest caveat
“Substantially faster” is a company describing its own system. No published benchmarks, no numbers next to the claim, no third-party verification. The merge volume growth is a concrete figure. The performance improvement is, for now, a characterization. If I were evaluating Scion as a product, that’s the first thing I’d ask for, and the absence of it isn’t damning but it does mean we’re taking the headline on faith.
It’s also internal. There’s no public release, no license, no community. The Hacker News thread when this landed was modest, a few dozen points and a handful of comments, which is a reasonable signal that the industry sees this as interesting rather than urgent.
Does this mean anything for your stack
Almost certainly not yet, and I want to be direct about that because the pattern-matching instinct in this industry is strong. Anthropic solved a problem it created by virtue of being an AI lab that runs agents against its own monorepo at a scale essentially nobody else is operating at. If your team merges 40 changes a day, Git is not your constraint, and migrating off it would be a self-inflicted wound.
The useful signal isn’t the tool. It’s the direction. The sequence of events here is worth tracking because it suggests where cost moves as agentic coding gets more common:
- Code generation gets cheap, so volume climbs.
- Review, merge, build, and test become the new queue.
- Tooling built around human working rhythms starts showing its assumptions.
You can watch for that in your own setup long before you’d need a custom VCS. Rising CI wait times, checkout costs that scale badly as the repo grows, merge queues that never quite drain. Those are the early versions of the same problem, and they’re fixable with sparse checkouts, better caching, and smarter CI scoping using tools that already exist.
My read
Scion is a credible response to a real and well-documented pressure, built by a team with an unusually clear view of where agentic development is headed. It’s also unverifiable from the outside and irrelevant to the vast majority of engineering teams for the foreseeable future.
Treat it as a weather report rather than a tool recommendation. Someone had to hit the wall first, and the specifics of how they hit it are more valuable than the custom system they built to get past it.
đź•’ Published: