Experiments in Scaling Autonomous Coding Agents at Cursor

Summary

Wilson Lin at Cursor conducted experiments with large-scale autonomous coding agents, running hundreds of concurrent agents on a single project that collectively wrote over a million lines of code and processed trillions of tokens. The system used a hierarchical approach with planners and sub-planners to create tasks, workers to execute them, and judge agents to evaluate completion. The post shares insights on scaling autonomous coding systems and coordinating large fleets of AI agents for software development.