swarm scaling

Improving performance on a task by increasing the number of AI agents working on it in parallel while holding each agent's reasoning length fixed — a form of inference scaling alongside longer chains of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate. (see test-time computetest-time compute, n.: The computing effort an AI model spends while answering a question, as opposed to while being trained. Letting a model think longer or try many answers can make it markedly smarter.). Analyzing published multi-agent results for frontier models in 2026, Toby Ord found a single agent the most token-efficient: each fourfold increase in agents roughly doubled the total tokens needed for the same score, so ten agents working together matched a single agent given only about three to five times as many tokens. The payoff is speed, since parallel agents can finish in a fraction of the wall-clock time. Ord expressed the coordination penalty with economists' 'stepping on toes' parameter λ — N agents together match one agent working N^λ times as long — estimating λ at roughly 0.5–0.7 across benchmarks, values that matter for models of an intelligence explosionintelligence explosion, n.: The idea that once AI gets good enough at improving AI, each smarter version could build a smarter one faster still, so abilities would snowball far beyond human level in a short time..

Swarm scaling got us the answer overnight instead of in a week, at four times the token bill.