Picture a small-town library with one checkout desk and one librarian stamping cards. It works fine for decades. Then the patrons change. They don’t browse, they don’t sleep, and each one wants the entire collection handed over before it starts reading. Twenty-five million times a day.
That is roughly the position Anthropic described in its October 13, 2026 engineering post about replacing Git internally with an in-house system called Scion. The numbers are the story. In May 2026, the company averaged around 2,500 code changes merged per day. By September, that figure was north of 30,000. Same engineers, same org, twelve-fold jump in four months. The merge queue stopped being a queue and became a firehose.
The part that makes this more than a vanity project
Plenty of large companies have built custom version control. Google did it, Meta did it, and the usual reason is scale of repository size or headcount. Anthropic’s reported trigger is different, and that is what makes it interesting for anyone who reviews tools for a living.
The constraint was not more people. It was agents doing the work people used to do, at a rate no human team would generate. Anthropic’s own framing is that it became bottlenecked by developer infrastructure built for human scale rather than agent scale. That is a sentence worth rereading if you are currently wiring coding agents into your CI pipeline.
The supporting figures: a monorepo carrying close to 100 GB of main-branch history, and continuous integration performing roughly 25 million checkouts per day. Git’s model of handing a client a full object graph is elegant when a developer clones once and works locally for a week. It is punishing when your consumers are ephemeral jobs that each want a fresh tree, finish in seconds, and disappear.
What Scion reportedly buys them
Two claims stand out from the post. First, Scion’s infrastructure footprint is 200x smaller than Git’s in CPU and memory terms. Second, CI and research jobs check out code two orders of magnitude faster. A hundredfold speedup on an operation you run 25 million times a day is the kind of math that justifies a rewrite, and reporting on the post describes a five-month migration to get there.
I want to be careful here, because I have not run Scion and neither has anyone outside Anthropic. These are vendor-reported numbers from a company describing its own work, measured against its own workload. That does not make them wrong. It does mean the 200x figure is specific to a 100 GB monorepo under agent-driven load, not a universal constant you can quote in your next architecture review.
The honest take for the rest of us
My read: this is a useful signal and a terrible template.
Useful signal, because it points at a failure mode most teams have not hit yet but will. If you add agents to your workflow, the first thing to break is rarely the model. It is the plumbing around it. Checkout time, merge queue throughput, test runner concurrency, artifact storage, log volume. All of it was sized for the rhythm of people who commit a handful of times a day and go to lunch.
Terrible template, because almost nobody should respond by writing their own version control system. Anthropic has a 100 GB monorepo, 30,000 daily merges, and the engineering depth to spend five months on infrastructure that ships no product features. If your repo is 2 GB and you merge 40 changes a day, your answer is shallow clones, better caching, and a larger CI tier. Not a greenfield VCS.
The practical version of this lesson is cheaper. Instrument the boring layer. Track how long your agents spend waiting on checkouts versus thinking. Watch whether merge conflicts scale faster than your throughput. Those are the metrics that tell you when human-scale assumptions start costing you, and they are measurable long before you need to replace Git.
Where this goes next
There is a hint that the thinking is spreading beyond one company. GitLab has been publicly discussing next-generation source code management with Anthropic, including a session at GitLab Transcend 2026 featuring Kranthi Erusu from its platform engineering side. That suggests the vendors building version control for everyone else are taking the agent-scale problem seriously rather than waiting for customers to complain.
That is the outcome worth hoping for. Most teams cannot build Scion, but most teams will eventually need something that behaves like it. If the pressure Anthropic hit at 30,000 merges a day pushes commercial tooling toward cheaper checkouts and smarter server-side operations, the benefit reaches the people who never read the engineering post.
For now, treat Scion as a weather report, not a product recommendation. Something is coming for your build infrastructure. You have time to prepare, and preparing looks a lot less dramatic than rewriting Git.
🕒 Published: