Cursor, GitLab, and Zed race to rebuild developer infrastructure for the AI era

GitHub's dominance in code hosting is facing its first real challenge in over a decade—not from a direct competitor, but from the shift toward AI agents handling much of the development work.

With SpaceX backing, Cursor announced Origin this week: a Git-compatible platform built for a world where agents, not humans, write most code. Meanwhile, GitLab unveiled Project Switch (its next-generation source code management backend), and Zed introduced DeltaDB, which ditches Git's commit model entirely for a continuous stream of deltas tied to agent conversations.

The infrastructure crisis#

GitHub now processes 1.4 billion commits per month—agents alone generating over 17 million pull requests monthly. That volume has created hundreds of incidents over the past year, exposing a fundamental mismatch: the pull request, a feature designed for human code review in 2008, doesn't scale for machine-driven workflows.

As Brian Douglas, former GitHub director of developer advocacy and now founder of Paper Compute, told The New Stack: "Agents are quickly killing the will for doing open source." Power users increasingly do code review and PR work directly in AI tools rather than returning to GitHub.

A new playbook for developer tools#

The pattern mirrors what happened with developer sandboxes: Docker, Cloudflare, and Vercel moved into that space because developers were already spending time there. Now IDEs are shifting from "places to type" to "places to watch agents write." Version control follows the same gravity.

Origin remains in waitlist mode with a planned fall launch. GitLab's Project Switch claims 50x faster agent task execution with 3x fewer tokens consumed (Anthropic is a design partner). Zed's DeltaDB goes further: instead of commits, it logs fine-grained operations tied directly to the conversations that produced them.

The model is the moat#

What ties these efforts together is ownership. Cursor launched its own coding model, Composer, and iterated to version 2.5—achieving a tenfold cost advantage on output tokens versus Claude Opus through in-house inference. As Douglas puts it: "You need to own the model to win." Relying on API calls from OpenAI or Anthropic, the old playbook, no longer scales.

Mitchell Hashimoto, HashiCorp co-founder and early predictor of this shift, tweeted in December: He's now an investor in (ERSC), another Git-compatible platform designed to handle thousands of commits per second.