What jumps out to me is not that Anthropic built something new, but that Git, as usually used, seems to have become the bottleneck for a workload that is now basically machine-generated churn. Thirty thousand merged changes a day is the kind of number that makes a lot of “normal developer workflow” assumptions look quaint. At that point, the system you use for source control stops being a pretty interface for humans and starts being a data pipeline.
I’m actually more interested in the parts Anthropic says they gave up than the parts they gained. Moving commits, merges, rebases, blame, and history queries onto the server side sounds very sensible if your clients are mostly agents and CI jobs. The trade-off is obvious too: less local independence, less of the classic Git “I can do this offline in a hotel room” story. For a company whose repo seems to live inside a highly managed internal environment, that’s probably acceptable. For everyone else, maybe not. That distinction matters, and Anthropic is right to say this is not a general-purpose Git replacement.
The piece that feels most transferable is the server-side API idea. A lot of Git pain comes from pretending every operation has to be expressed as a checkout plus local manipulation of a giant working tree. If most of your real work is “give me the diff,” “give me the files touched by this changeset,” or “let me run a bot over this subset,” then the local clone is just an expensive compatibility layer. I think that’s the real lesson here: not “replace Git,” but “stop assuming local clones are the center of gravity.”
That said, I’m a little cautious about how much of this is architecture and how much is just Anthropic being Anthropic. The company reports huge cache hit rates and dramatic footprint reductions, but those are internal numbers for an internal workload. They may be real and still not tell you much about whether this would work outside a tightly controlled monorepo with bespoke code review, code search, and CI. The article hints at that, and I think that warning should be taken seriously.
Still, the direction feels right. Once AI agents and CI are doing the bulk of repo interactions, content-addressed storage, tiered caches, and server-side derived data start to look less like clever optimizations and more like the natural shape of the problem. Git wasn’t designed for a world where machines interrogate the repo constantly. Anthropic seems to be building for that world instead of arguing with it.
Reference: How Anthropic Redesigned Source Control for AI-Driven Development
