Cursor Builds Its Own AI Model and a Git for Agents

Cursor announced three new products on June 23: its first fully self-developed AI model, a new code-hosting platform called Origin, and an iOS beta for Cursor Mobile. Each reflects a different piece of the same larger bet — infrastructure for a world where AI agents write most of the code.

The model is the biggest shift. Cursor has always routed work through third-party models — Claude, GPT, and others — and competed primarily on the quality of its editor and how it manages context. The new model, comparable in scale to Opus and GPT-class models, was trained with 10 to 20 times more compute than any model Cursor has used before. It was built to go beyond coding: the company describes it as designed for work that spans multiple application types. It uses the SpaceX compute infrastructure from the recent acquisition and ships within weeks.

Why AI agents break standard Git

Origin solves a problem that has been building quietly. Standard Git was designed for humans working at human pace — a few dozen contributors, code reviews happening over hours. When hundreds or thousands of AI agents all write to the same codebase simultaneously, the system collapses: edit conflicts pile up faster than they can be resolved, automated tests fail in batches, and the feedback loop that a human reviewer provides grinds the whole thing to a halt. Cursor tested Origin at exactly that scale — thousands of agents writing in parallel — and built in automatic conflict resolution, test repair, and the ability to handle reviewer comments on agent-generated code. It is currently live inside Cursor and with select external partners; broad availability is planned for fall 2026.

The mobile app closes the supervision gap. Cursor agents already run multi-hour sessions — writing code, running tests, iterating through failures without anyone watching. The iOS beta lets you monitor progress from wherever you are: review screenshots the agent generated, unblock it when it hits a decision point it can't resolve alone, or remotely control agents running on your local machine. The long coding session that used to keep you at your desk can now run while you step away.

What this triple announcement signals

These three products form a coherent stack: the model at the core, Origin as the coordination layer for the agents that model will power, and mobile as the interface for the humans still in the loop. It's an unusual package for a company that started as a code editor. Cursor is now building the full pipeline for software development at agent scale — including the infrastructure pieces that were never designed for it.