Artificial Intelligence & Computing

How AI Agent Swarms Are Rewriting Model Economics

For years, getting a smarter AI meant buying a bigger, more expensive model. A new approach flips the equation: spread the work across hundreds of small agents and reach frontier quality at a fraction of the cost.

• A swarm of hundreds of cooperating AI agents just beat a single elite model on a notoriously hard software task, passing 80 percent of a held-out test suite in four hours.
• The trick is not raw parallelism. The swarm splits each job into a planning layer and a working layer, so no agent ever has to carry both the big picture and the messy details at once.
• By mixing a strong model for planning with cheap models for execution, the same result can be produced at dramatically different prices — opening a genuinely new cost curve for AI.

For the last few years, progress in artificial intelligence has largely followed one rule: bigger models get better answers, and bigger models cost more to run. Reasoning models turned that dial up further, spending enormous amounts of compute on each individual question. The natural assumption was that the only path to harder tasks was simply to throw more tokens and dollars at the problem.

A new experiment from the team behind Cursor suggests a different path. They built an AI "swarm": hundreds of autonomous agents that share a single task, rather than one giant model taking it all on its own. To test it, they gave the swarm an unforgiving job — rebuild SQLite from scratch in Rust, using only the public documentation, with no existing code to lean on. The new swarm reached 80 percent on a held-out SQL test suite in four hours. The older single-agent approach spiraled and had to be stopped before its second hour.

The advantage comes from how work is organised. Every large task can be drawn as a tree: a goal at the top, splitting into sub-goals, splitting again, until each branch ends in a small, concrete action. In a swarm, a planner stays near the top of the tree and never gets bogged down in low-level detail. A worker stays at one leaf and never has to hold the whole project in its working memory. Each agent gets a context size that fits the size of the job it is actually doing.

The headline in one sentence: coordinating hundreds of modest agents around a shared plan can match a single frontier model, while using far less of its most expensive compute.

That opens a financial implication almost as important as the technical one. Because planning and execution are separated, teams can run the strategy on an expensive reasoning model and hand the routine work to a fast, cheap model. Cursor's runs showed that nearly every model mix produced similar quality, but the costs varied enormously. In other words, the same answer can be bought at very different prices depending on how the swarm is organised.

The swarm also had to solve problems that no human engineering team ever sees. At a rate of roughly a thousand code commits per second, ordinary version control collapses, and two planners can unknowingly decide the same architectural question in opposite ways. The team answered with new internal tools: a version-control layer built from scratch, shared design documents that propagate decisions downstream, and neutral "referee" agents that resolve conflicts the working agents cannot. These are not just engineering fixes — they are the first clues about how coordination economics will scale as swarms grow from hundreds to millions of agents.

The broader lesson is familiar from economics itself. The economist Ronald Coase once asked why firms exist at all, and answered that when too many people try to coordinate directly, the overhead grows faster than the work — so organisations settle into tiers of small units. AI swarms, it turns out, obey the same rule. The next wave of capable AI may come not from a single ever-larger brain, but from a well-organised team of modest ones.