Agent swarms and the new model economics

Cursor's blog post describes their agent swarm architecture, the engineering challenges of running it at scale, and what the economics say about using multiple models together.

The swarm uses a tree-like decomposition: planner agents (frontier models) split goals into subtasks, and worker agents (cheaper, faster models) execute them. Planners and workers never swap roles — planners don't implement, workers don't plan. This separation keeps each agent's context focused and is cited as the reason long-running single agents don't drift.

Cursor built a custom version control system to handle swarm throughput of ~1,000 commits per second. They catalog five failure modes at that scale: split-brain design (two planners deciding the same thing differently), planner contention, merge conflicts, megafile bloat, and ossification (agents afraid to touch core code). Each has a structural fix — for example, ossification is solved by licensing intentional breakage, with the compiler propagating changes downstream.

Review is done with stacked, decorrelated lenses (different models, perspectives, scopes) — no single lens catches everything, but they compound. Review compute is cheap relative to the work it audits.

The swarm uses stigmergy: agents coordinate by shaping a shared environment. A self-authored Field Guide folder, curated by agents, is injected into every new agent at startup to capture surprising encounters worth sharing.

The centerpiece experiment is implementing the SQLite spec (835 pages) in Rust without access to source code, test suites, or the internet, graded against sqllogictest. A Fable 5 planner with Composer 2.5 workers passed ~2/3 of the suite in the first hour; every model mix eventually reached 100%. The new harness produced far fewer commits, merge conflicts, and lines of code than the old one — old runs thrashed, new runs converged.

On economics: running GPT-5.5 alone cost $10,565 (workers burned $9,373). The Opus 4.8 + Composer 2.5 hybrid cost $1,339, with the entire worker fleet at $411. Workers carried ≥69% of tokens in all runs, but planner tokens dominated cost. The thesis: few moments in a large task genuinely require frontier intelligence, so split the work accordingly.

The post frames swarms as the next abstraction jump — from autocomplete to lines to blocks to files to specs. A spec becomes the prompt; the swarm is a probabilistic compiler translating intent into working code.