Agent swarms and the new model economics

Wilson Lin documents Cursor's rebuilt agent swarm architecture, where planner agents on frontier models decompose a goal into subtasks and worker agents on cheaper models execute them. The stress test was rebuilding SQLite from its documentation: the new swarm hit an 80% test pass rate in four hours while the old system spiraled within two, drowning in tens of thousands of merge conflicts at a thousand commits per second. Two findings stand out. Hybrid configurations reached similar quality at wildly different costs, and the whole exercise relocates the bottleneck: when execution is this cheap, the unit of work becomes the spec, and the scarce resource is the right description of intent.