Picture a Tuesday afternoon in your own pipeline. You’ve got six agents running in parallel on a single repo, each one cloning, branching, committing, and opening a pull request. Your CI box starts wheezing. The checkout step, the part you never think about, the part that’s supposed to be free, is now the longest bar in the trace. You stare at the timing breakdown and realize your agents are spending more wall-clock time waiting on version control than writing code.
Now multiply that by a few thousand. That’s roughly the wall Anthropic hit, and on 13 October 2026 the company published an Engineering blog post explaining what it did about it: built an in-house source control system called Scion and moved off Git internally.
Numbers that explain the decision
The growth curve is the whole story. In May 2026, Anthropic averaged around 2,500 code changes merged per day. By September, that number was over 30,000 per day. Twelve times the merge volume in four months, and the company said it doesn’t expect the growth to slow.
The supporting figures are just as telling. The monorepo carried nearly 100 GB of main-branch history. CI performed about 25 million checkouts per day. If you’ve ever watched a Git clone crawl through a large history, you can do the arithmetic on what 25 million of those looks like as a daily infrastructure bill.
The results they reported after the migration, described elsewhere as a five-month effort:
- Scion’s infrastructure footprint is 200x smaller than Git’s in CPU and memory requirements
- CI and research jobs check out code two orders of magnitude faster
Two hundred times less compute for the same job. That’s not a tuning win. That’s a sign the original tool was being asked to do something it was never shaped for.
Human scale versus agent scale
Anthropic’s own framing in its announcement is the part I keep coming back to: they became bottlenecked by developer infrastructure built for human rather than agent scale. That sentence should be pinned above every bot builder’s desk.
Git’s design assumptions are reasonable assumptions about people. A developer clones once and keeps the working copy around for months. Branches live for days or weeks. History is something you read occasionally, usually when something has gone wrong. Merge conflicts happen a handful of times a week and a human resolves them with context and judgment.
Agents break every one of those assumptions. They clone constantly because they’re ephemeral. They create and discard branches in minutes. They don’t need decades of history to make a two-line change, but they pay for it on every checkout anyway. And they generate merges at a rate no review process designed around human attention was ever sized for.
None of this makes Git bad. It makes Git a tool whose cost model stopped matching the workload.
What this means if you’re not Anthropic
Let me be clear about the thing I am not saying. You should not go write your own version control system. Anthropic has a specific scale problem, specific internal resources, and a monorepo shaped by years of decisions. Copying the solution without the problem is how teams end up maintaining infrastructure they don’t need.
But the diagnosis generalizes beautifully, and that’s the useful part. If you’re building agent systems, the lesson is this: your bottleneck will probably show up in the plumbing, not the model.
Things worth measuring in your own setup
- Checkout time as a share of total job time. If it’s climbing, you have a Scion-shaped problem in miniature.
- How much history your agents actually need. Shallow clones, partial clones, and sparse checkouts exist and are badly underused. Most agent tasks need the current tree, not the full ancestry.
- Merge throughput versus review throughput. Generating 30,000 changes means nothing if your gate downstream handles 300.
- Per-task infrastructure cost. A 200x difference in CPU and memory on one step is the kind of thing nobody notices until it’s the entire bill.
There’s also a signal in the broader ecosystem. GitLab and Anthropic have been discussing version control for large-scale AI development publicly, which suggests the tooling vendors are already thinking about this. If you’d rather not build, waiting is a legitimate strategy.
The architectural read
What I find genuinely interesting here isn’t the new system. It’s the admission embedded in building one. Anthropic looked at a tool that works fine for essentially the entire industry, measured it against its own workload, and concluded the mismatch was structural rather than fixable through configuration.
That kind of clarity is rare, and it’s the skill worth copying. Most of us will never need to replace Git. All of us will eventually hit something in our stack that was built for a different kind of user. Noticing it early, with numbers rather than vibes, is the whole job.
Go look at your checkout times. Seriously. It’ll take five minutes and it might surprise you.
🕒 Published: