When should you use multiple AI agents instead of one?

Anthropic's research system uses a lead agent that spawns subagents to search in parallel. On its internal evaluation the multi-agent version outperformed a single agent by 90.2%. The same write-up is clear about the price: agents used about four times the tokens of chat, multi-agent systems about fifteen times, and the approach is less suited to tasks where agents must share context or depend on each other — it names coding specifically.

Cognition, which builds coding agents, argues the other side in “Don't Build Multi-Agents”: agents should share full context, and parallel agents make implicit decisions that conflict. Its recommendation is a single-threaded agent with context compression for long tasks.

OpenAI's guide lands in between: maximise a single agent's capabilities first, and split when the logic becomes too complex or the tools too many and too similar for one agent to choose between.

Cognition illustrates the risk with a Flappy Bird clone split across two subagents: one builds the background, the other the bird, and each makes reasonable stylistic choices that do not match. The combined result is inconsistent, even though each part is fine on its own. Research is different: two subagents searching different sources rarely contradict each other in a way that breaks the result.

Multiple agents can cover more ground at once and keep each agent's context focused. They cost more, add coordination failures — which the MAST study catalogues across 1,600 traces — and make debugging harder because a failure may come from the hand-off rather than any one agent.