Cursor's Agent Swarm Experiment: Same Result, 15x Cost Swing Depending on Model Mix

Cursor·high signal

Cursor published research (Wilson Lin, July 20) on scaling cooperating agent swarms, in which a team of agents rebuilt SQLite in Rust from scratch using only its 835-page manual — no source code, no test suites, no internet — and passed a held-out sqllogictest suite. The economics are the headline: the cheapest model configuration cost $1,339 while running everything on GPT-5.5 cost $10,565 for the same result. Frontier models were reserved for high-judgment moments while cheaper models did the bulk execution once the plan was fixed — a directly copyable pattern for anyone running multi-agent pipelines on a budget.