CURSOR'S AGENT SWARM SHOWS CHEAP MODELS CAN CODE

Cursor tested its new agent swarm architecture against the old one in a challenging task: rebuilding SQLite in Rust using only documentation, with no source code or internet access. The upgraded system separates roles between planner and worker agents. Frontier models handle planning and high-level decisions, while cheaper models execute the actual coding work. Every configuration of this hybrid approach achieved 100 percent accuracy on the test suite. The predecessor system, which lacked this hierarchical structure, struggled significantly. It became trapped in merge conflicts created by its own changes, unable to resolve them without human intervention. The results suggest a cost-effective path forward for AI-assisted development. By reserving expensive frontier models for complex reasoning and delegating execution to cheaper alternatives, developers could reduce inference costs while maintaining quality output. The architecture appears particularly effective for multi-step programming tasks that require both strategic planning and implementation.