Swarm Scaling Quantifies Agent Coordination Tax

Swarms look less like a free intelligence multiplier and more like an expensive speed dial. Their value depends on whether faster results justify the coordination and compute premium.

• –Four-agent swarms may reach similar performance with roughly twice the total tokens, while cutting wall-clock time by about half.
• –A 10× increase in agents produces only about 3–5× the performance gain under Ord’s estimates.
• –Swarms are most compelling for parallelizable tasks with automated verification, such as coding, mathematical search, and simulations.
• –Identical agents can amplify shared mistakes; diverse agents and independent checking may reduce correlated failures. [Understanding AI analysis](https://www.understandingai.org/p/why-agent-swarms-could-be-the-next)
• –Developers should optimize for marginal quality per token and latency, rather than assuming more agents automatically means better answers.