Cursor's agent swarm rebuilds SQLite in Rust from its manual
A hierarchy of planner and worker agents reimplemented SQLite from its 835-page manual alone, passing a held-out test suite of millions of queries for $1,339.
Cursor has published one of the most striking agent-autonomy demonstrations to date: a swarm of AI agents reimplemented SQLite from scratch in Rust, working only from the database's 835-page manual — no source code, no test suites, no internet access. The resulting replica passed 100% of a held-out sqllogictest benchmark containing millions of queries with verified answers.
Planners think, workers type
The experiment, detailed in a July 20 blog post on agent swarms and model economics, hinges on hierarchical decomposition. Planner agents running frontier models break the goal into a tree of tasks; worker agents running faster, cheaper models execute the pieces. Cursor says this division fixed the failure modes that killed earlier multi-agent runs — context bloat, merge conflicts between agents, and "split-brain" divergence — because planners never touch implementation details and workers only see their narrow slice.
The cost data is as notable as the result. The cheapest successful configuration, pairing Claude Opus 4.8 as planner with Cursor's Composer 2.5 as workers, finished for $1,339. Running everything on GPT-5.5 produced the same outcome for $10,565 — roughly eight times more. With Grok 4.5 in both roles, the new swarm architecture reached 80% test completion within four hours, while Cursor's previous system had already spiraled before hour two. The company's takeaway: only a few moments in a large task genuinely require frontier intelligence, and cheaper models can follow a good plan.
Why it matters
Rebuilding a well-documented database is not the same as novel engineering, and skeptics were quick to note that SQLite's spec-and-test ecosystem makes it an unusually clean target. But the demonstration lands two real points: multi-week software projects can now be decomposed and completed by coordinated agents with minimal human input, and orchestration choices — not just model choice — drive an order-of-magnitude cost difference. For teams budgeting agentic development, the planner-worker split is quickly becoming the reference architecture.