Anthropic Left Git. The Bottleneck Didn’t Leave With It.
Anthropic opened its engineering post with two numbers:
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Thirty thousand merges a day is one every three seconds, around the clock. The post, Redesigning source control at Anthropic, describes how seven people built a new version control system called Scion and moved the company's monorepo off git in five months. Most of the coverage stopped at "Anthropic ditched git". I think the more interesting story is about what broke, why, and where the pressure goes next.
What expired was git's assumption, not git
Git was designed for a world where every developer holds a full copy of the repository and can work offline for a week. That is a good design for people. It is a bad one for a client that never goes offline, lives for ten minutes and needs three files.
That is what Anthropic's clients are now: agents and CI jobs. The main branch history is almost 100 GB, and CI alone checks out code about 25 million times a day. Anthropic estimates that fetching naively would mean more than 600 PB of git traffic a day. They tried every standard workaround first: shallow clones, sparse checkouts, promisor remotes, snapshots in blob storage and, at the peak, several hundred internal git mirrors. Each one patched the cost of every client carrying the repo. None of them removed it.
Scion removes it. The server keeps the history and runs blame, merges, rebases and commits. Clients ask for the trees and blobs they actually read, through a stack of caches, and never send a list of what they already have.
The most common joke about the post is that Anthropic reinvented Subversion. It is closer to true than the people making it think. A central server, thin clients, commits that cost the size of the change rather than the size of the repo: that is the old model, and Google's Piper and Meta's Sapling went the same way years ago. The pendulum swung to decentralised version control because developers were often offline. Agents never are, so it is swinging back.
Every layer built for human pace breaks, one after another
Anthropic is not the only one hitting this. In August GitHub had an eight-hour outage, and GitHub itself said AI-driven growth had overwhelmed its infrastructure. Cursor launched Origin, its own code hosting built around agents as the main readers and writers, on the same day.
Source control is just the layer that broke first, because it sits right next to the code. Any system that assumes a person sits between two steps is a queue waiting to overflow: a merge queue that takes a global lock, an API rate limit sized for a human clicking, an approval step that expects a few PRs a day. Anthropic dropped the global-lock merge queue entirely. It now relies on testing after merge and before deploy, and lets Claude fix conflicts, which it usually does faster than a person would notice them.
The bottleneck moves closer to a person
The obvious objection to the post writes itself: nobody is reading 30,000 diffs a day. Anthropic agrees. Its engineers now review the intent of a change rather than every line, and the post says plainly that their main job was often to tell Claude "no" when a solution broke a system constraint.
That answer works, and it is also the next bottleneck. I wrote in August that code got cheap but understanding didn't. Scion is that same story one floor down. Making merges cheap doesn't make review cheap. It moves the narrowest point one step closer to a human.
The same pattern shows up outside source control. The New Stack reported that Claude flagged about 29,000 suspected bugs in open source and only 516 of them were fixed. Finding things scales with compute. Accepting them still goes through a maintainer.
What the reactions get right and wrong
| Claim | Evidence | Verdict |
|---|---|---|
| "Git is obsolete" | Anthropic says Scion is not a general git replacement. It matters for very large monorepos, and most repos are nowhere near 100 GB. | Overstated |
| "They reinvented SVN" | Central server, thin clients, server-side commits. | Basically right, and that is the point |
| "Seven people did this in five months, so small teams win" | True as reported, but those seven had Anthropic's internal models and compute. One internal model rewrote the ingestion pipeline to use 20x less memory and run 4x faster. | Real, hard to copy |
| "They won't share it, so everyone has to rebuild it" | Not open source. Anthropic says it wants to bring the ideas to git, GitHub and GitLab instead. | Half fair |
What I take from it
Look for the assumption, not the tool. Somewhere in your stack there is a step that quietly assumes a person goes between two actions at human speed. That is where agents will hit first.
Don't count on review scaling. Anthropic's own answer is intent review plus tests after merge, and it still depends on someone who can say "no" with good reason. That person is now the scarcest part of the pipeline.
For most teams git is fine. Borrow the ideas instead: partial clones, sparse checkouts, no global locks that agents have to wait on.
A canal lock lets one boat through at a time, and that worked as long as the boats came at the pace of the horses towing them. Anthropic widened the channel and took the gates out. Further down the river, someone still has to decide which boats should have come through at all.
Header image: John Constable, "Flatford Mill from the Lock" (c. 1810), Clark Art Institute, public domain.