Cursor's agent swarm rebuilt SQLite in Rust with 100% test pass rate, proving cheap models can ship

Cursor’s latest agent swarm didn’t just beat its predecessor—it embarrassed it. The task was brutal: rebuild SQLite in Rust using only the 835-page documentation. No source code, no test suites, no internet. Every configuration of the new system eventually hit 100 percent on the hidden sqllogictest suite. The old swarm? It drowned in its own merge conflicts.

The breakthrough is a strict division of labor. Planner agents use expensive frontier models like Opus 4.8 or Fable 5 to recursively break goals into smaller tasks. Worker agents, running cheaper models like Composer 2.5, just execute. The planners never write code. The workers never plan. This solves the context drift that kills long solo-agent runs.

Speed exposed new problems. The old system managed about 1,000 commits per hour. The new one hit 1,000 per second, breaking Git entirely. Cursor had to build custom version control. They also had to solve a “split-brain” problem where two planners would unknowingly build the same feature in different ways. The fix involved shared design documents with compile-time reference checks and neutral agents that resolve merge conflicts.

The old Grok 4.5 run is a case study in wasted effort. It produced 68,000 commits in two hours—70 times more than the new system—and racked up over 70,000 merge conflicts. The most contested file saw 7,771 conflicts from 1,173 agents. The new run stayed below 1,000 conflicts total. When Fable 5 planned the work, the old swarm needed 64,305 lines of engine code. The new one did it in 9,908 lines and scored 100 percent. The cost savings are real: total runs ranged from $1,339 to $10,565, and since worker models handle at least 69 percent of tokens in every run, using cheaper models for execution is the biggest lever.