Agent swarms and the new model economics

Cursor experimented with an Agent Swarm structure in which agents collaborate to achieve complex goals. Unlike the existing single-agent approach, which suffers from a 'drift' phenomenon of losing context and losing sight of the goal during long task execution, the swarm structure solves this by decomposing tasks into a hierarchical tree structure.

This system performs two key roles:

• Planner agents: Use a top-tier model to split the goal into subtasks and delegate them.
• Worker agents: Use fast, low-cost models to execute the assigned detailed tasks.

As a result of the experiment, the new swarm structure showed outstanding performance on the highly challenging task of rebuilding SQLite in Rust from scratch. Using Grok 4.5, on a task where the existing swarm failed to last even 2 hours, the new swarm succeeded, passing 80% of the SQL test suite in 4 hours.

In addition, meaningful results were also obtained in terms of economics depending on model combinations. Instead of using a high-performance model for all tasks, a structure in which a high-performance model handles planning while a low-cost model handles execution proved that it is possible to dramatically reduce costs while maintaining quality.