Smart AI planning makes coding faster and cheaper

The team at Cursor tested their new agent system by asking it to rebuild SQLite in Rust. They gave the AI only the documentation to work with. They did not allow the agents to use the internet or existing code.

The new setup works by separating the planners from the workers. A smart model plans the steps, while smaller and cheaper models carry out the tasks. This approach reached a perfect score on every test.

Older systems struggled with this task because they created errors they could not fix. By splitting the workload, this new method proves that big models are not always needed for every step of the coding process.