Cursor has spent weeks testing autonomous coding agents on massive projects, pushing the frontier of agentic development. They ran hundreds of concurrent agents on single codebases, generating over a million lines of code and trillions of tokens. The experiments aimed to tackle projects that typically demand months of human effort, revealing both the potential and the pitfalls of scaling AI beyond simple, focused tasks.

Initial attempts at coordination failed. Letting agents self-organize through shared files and locks created bottlenecks, with twenty agents slowing to the throughput of just two or three. Without hierarchy, agents became risk-averse, avoiding difficult problems and churning on small, safe changes. This left complex end-to-end implementation work untouched, proving a flat structure couldn't handle ambitious goals.

The breakthrough came from dividing labor. Cursor created a pipeline with Planners exploring the codebase and creating tasks, while Workers focused solely on execution. This structure solved coordination problems, allowing them to scale effectively. They tested it by building a web browser from scratch and migrating Solid to React in their own codebase, proving hundreds of agents can work together for weeks with minimal conflicts.

Model choice proved critical for these long-running tasks. GPT-5.2 models outperformed others at maintaining focus and following instructions over extended periods. Cursor also learned that simpler systems often work best; adding an integrator role created more problems than it solved. Ultimately, careful prompting matters more than complex architecture. The goal is to integrate these findings into Cursor's product.