Will AI agents push teams off git?
Anthropic replaced git with its own version control system after agents pushed its merges from 2,500 to 30,000 a day. Polly Morphic says every busy team is next; Cassandra Prompt says big companies hit this wall years ago, and git kept up.
Opinion. The views are Polly Morphic's and Cassandra Prompt's own and are not investment advice.
This week Anthropic said it had moved its main code repository off git and onto Scion, a version control system it built itself. The reason, in its engineering post, is volume: with Claude writing more and more of its code, merged changes went from about 2,500 a day in May to more than 30,000 a day in September, and its CI checks out the repository about 25 million times a day. Git gives every machine a complete copy of the repository; Scion keeps the repository on its servers, which do the heavy work, and each agent or job fetches only the files it needs. Anthropic expects “this class of problems” to become common within six to 12 months. If your agents open more pull requests every month, the question is whether that warning is about you.
Polly Morphic: git was built for people, and people are leaving the loop
Where I file from, in 2031, git is still everywhere, the way FTP was still everywhere in 2010. It works. It’s just not where the interesting work happens.
Look at what broke at Anthropic. Not the code; the code got written. What broke was everything that assumed a human behind each change: a design where every machine keeps its own copy of the repository (Anthropic had already patched that with partial clones and hundreds of mirrors), a merge queue that lines changes up one at a time, CI that downloads the repository before it does anything. Those were sensible choices when a busy team merged a few hundred changes a day. They stop being sensible when the authors don’t sleep.
And the authors are multiplying. Anthropic’s merges went up twelvefold in four months. You don’t need to be Anthropic for that curve to reach you; you need a team that gives each engineer five agents instead of one. When that happens, the slow part is no longer writing code but getting it tested and merged, and your tools were built for the old slow part.
The detail I’d underline is the merge queue. Anthropic says it couldn’t build a queue that locks the repository for each change at this rate, so it now merges first, tests after merging and before deploying, and has Claude fix changes that clash once they’ve landed. That is not a scaling trick. It’s a different idea of what the main branch is: a stream that’s always moving and always being checked, not a gate you queue for.
And switching got cheap: seven people built Scion and a new code review tool and moved the whole repository in five months. “Everyone uses git because switching is too expensive” is no longer a law of nature.
So no, you won’t install Scion. But by the time your agents outnumber your engineers, GitHub or GitLab will be doing on their servers the work your copy of git does on your laptop today: history, diffs, merges. Git will be how your files are stored, not how your team works.
Cassandra Prompt: this is the monorepo problem again, and git kept up
I’ve seen this before. In 2017 Microsoft moved Windows onto git: about 3.5 million files, a 300 GB repository, around 4,000 engineers. Plain git couldn’t cope. Microsoft didn’t leave; it built a virtual file system so each machine downloaded only the files it used, and its successor, Scalar, built on git’s own partial clone and sparse checkout, has shipped inside git since version 2.38. Google and Meta built their own, Piper and Sapling, and Anthropic’s post credits both. Polly calls this the end of git. It’s one more very large company finding out that a very large repository needs a server that does the heavy lifting.
The post’s own caveats are the strongest part. Anthropic says it doesn’t expect many organisations to have these problems in the very near term. It says its ability to spend tokens internally “far exceeds what any other business would realistically want to spend”. It calls Scion “by no means a suitable replacement for git for general use”. That’s not a product launch. It’s a warning from the one customer at the far end of the curve.
That budget also answers Polly’s seven people in five months. They had early access to Anthropic’s newest models, a token budget no other business would want, and a problem big enough to justify both. Your team has none of the three.
On the merge queue, Polly sees a new idea of the main branch. I see a team that chose to let changes clash after they land and pay an agent to clean up. That’s a fair trade at their volume. At yours, a queue that tests before merging is what stops a broken main branch at 3am, and nobody should copy the removal of a safety check because a lab with a special budget had to.
Notice what Anthropic actually asks for: richer server-side APIs from GitHub and GitLab, which the ordinary git client could learn to use. That’s git adapting again, the way it did for Windows. If your CI is slow, try partial clones and sparse checkouts before you plan a migration. By Anthropic’s own account, they bought it time too.
Edgar Case rules
On Polly’s facts: the 2,500 and 30,000 merges a day, the seven-person team, the five months and the merge-first approach are all in Anthropic’s post, and “twelvefold in four months” is correct arithmetic. Five agents per engineer is her hypothetical; the post gives no such figure. The six-to-12-month warning covers “this class of problems”, not leaving git, and the same post says it doesn’t expect many organisations to hit them soon.
On Cassandra’s facts: the Windows numbers match Microsoft’s 2017 post, Scalar did ship in git 2.38, and Anthropic credits Piper and Meta’s Sapling. But her comparison is about size, and size alone is not what’s new: Windows was a 300 GB repository, while Anthropic’s main-branch history is about 100 GB. What’s new is the rate, 25 million CI checkouts a day, and Anthropic says it had already used shallow clones, sparse checkouts, partial clones and several hundred git mirrors before it gave up. Her advice suits most teams; Anthropic is the case where it ran out. The team did have an early internal version of a new Claude model, but tying the token budget to the five months is her reasoning: the post cites the budget to explain why Scion won’t suit other companies.
Ruling: Cassandra had the better of the facts on what the post says, Polly on what is actually new in it.
The predictions
Polly Morphic: “By 18 April 2027, a company other than Anthropic with more than 1,000 employees will publicly announce that it has moved, or is moving, its main code repository off git, so that its developers and agents no longer use git to check out and commit, and will give AI agents or agent-written code as a reason.”
Cassandra Prompt: “By 18 April 2027, no company other than Anthropic with more than 1,000 employees will have publicly announced moving its main code repository off git because of AI agents or agent-written code, and Scion will have no public release.”
Both go on the scorecard. Edgar settles them in April.