Cursor Analyzes Cost of Multi-Agent Swarms

Cursor is developing an agent swarm architecture that coordinates hundreds of AI agents, and it ran a benchmark implementing a Rust database engine from just an 835-page SQLite document.

Core design: planner-worker tree structure

• Planner: A top-performing frontier model hierarchically decomposes tasks and delegates them
• Worker: Cheaper, faster models execute individual subtasks
• Thanks to the tree structure, each agent keeps its context focused on a narrow role, mitigating the drift problem that occurs during long-running execution

Cost experiment results

• The new swarm outperformed the old swarm across every model combination
• Using Grok 4.5, it passed 80% of the SQL test suite in 4 hours (the old swarm stalled within 2 hours)
• At equal quality, cost ranged from $1,339 to $10,565 depending on the model combination (about an 8x difference)

Implications: The key conclusion is that on large tasks, frontier-model intelligence is actually needed only for a handful of moments—such as initial task decomposition and design decisions—while the rest of the execution can be handled by cheaper models. Cursor stated that this structure generalizes to a variety of tasks, including browser implementation, solving math problems, and GPU kernel optimization.