Cursor rebuilt SQLite with a swarm of agents, and the cheap models did most of it

Cursor set its agent swarm an unusually pure test: rebuild SQLite, the database engine that sits inside nearly every phone and browser, in the Rust programming language, using nothing but the 835-page manual. No source code, no existing test suite, no internet. The grading was done against sqllogictest, a suite of millions of SQL queries with known answers that the agents never knew existed. Every configuration of the new system eventually scored 100 percent.

The change is simple to describe. Agents are split into two roles. Planner agents, running expensive frontier models, break the goal into smaller and smaller tasks. Worker agents, running faster and cheaper models, carry those tasks out. Planners never write code, workers never plan. Cursor argues this is really a fix for a context problem: a single agent has to hold the overall goal and the current sub-task in mind at once, which is why long agent runs tend to drift. The results were lopsided. The old swarm produced 68,000 commits in two hours, about seventy times as many as the new one, and most of it was wasted: it piled up more than 70,000 merge conflicts, where the new run stayed under 1,000. Where the old system needed 64,305 lines of engine code, the new one needed 9,908.

The money is where this gets interesting for anyone who is not building a coding agent. Running the whole job on GPT-5.5 alone cost $10,565. Splitting it between Opus 4.8 as planner and Cursor’s own Composer 2.5 as worker cost $1,339, for comparable or better results. Workers consumed at least 69 percent of the tokens in every run, usually far more, but because worker models are cheap, they were not where the bill came from. In the Opus hybrid, the planner produced a small share of the tokens and two thirds of the cost. The lesson Cursor draws: only a few moments in a large task actually need frontier-level intelligence, mostly breaking down the work and making key design decisions. Once the ambiguity is resolved, cheaper models can follow the plan. A fair caveat: planner quality still mattered a lot. A Fable 5 planner used fewer planning tokens but left its workers needing far more, and the run cost more overall.

What this means for you: If you are just getting started with AI, the useful idea here has nothing to do with SQLite. It is that “which model should I use” is often the wrong question, because the better answer is usually “a smart one to plan, a cheap one to execute.” You can do a small version of this by hand: ask an expensive model to write you a clear step-by-step plan, then hand the individual steps to a cheaper one. If you build with agents professionally, Cursor published the codebase from one run as minisqlite on GitHub, and the planner-worker split is the part worth stealing. Keep expectations grounded, though: a 2025 study found that 68 percent of agents actually deployed in companies complete no more than ten steps before a human steps in. Lab swarms and production agents are still very different animals.

Sources

• The Decoder: Cursor’s agent swarm suggests cheaper models can handle most coding when frontier models plan the work
• Cursor: minisqlite on GitHub
• The Decoder: Composer 2.5 matches Opus 4.7 and GPT-5.5 benchmarks at a fraction of the cost