Version control after Git: what the AI era is actually changing

Git won twenty years ago and is not dying. What is changing — fast, in the last few months — is everything around it: who (or what) commits, how often, and where the intelligence of the system lives. Three signals from this autumn:

Anthropic outgrew Git internally. As coding agents flooded its monorepo, daily merges reportedly went from ~2,500 in May to over 30,000 in September 2026. Seven engineers spent five months replacing Git with an internal system, Scion, built on inverted assumptions: clients always online, the server does the computing, the working copy optional. Anthropic says Scion is not a general Git replacement — it is a pressure gauge.

Cursor folded hosting into the editor. Origin, in beta since August, keeps Git as the wire protocol but moves the forge — cloud agents, stacked PRs, merge queues, hosting — under one roof. The analysis making the rounds puts it well: Origin does not abandon Git; it abandons GitHub.

The client is being rebuilt for agents. Jujutsu (jj) treats the working copy as a commit, keeps an operation log, and makes conflicts first-class — and it runs on top of ordinary Git repos, so adoption costs nothing. Its creator now leads engineering at East River Source Control, a startup building the server side of the same idea. Smaller experiments, like the multiverse VCS Draft, ask what happens when ten agents share one repo at once.

My read, as someone who builds Git tooling for Unreal projects: the storage format survives, the client gets demoted to a compatibility layer, and the value moves to the server and the review surface — the layers that can answer not just what changed but who acted, why, and who is accountable. That is the design space worth building in now.