Cursor's experiment involved running hundreds of AI agents to build a web browser. The agents wrote over 1 million lines of code across 1,000 files in nearly a week. They used planners and sub-planners to create tasks for worker agents.