Multi-agent setups only beat solo agents on sparse tasks

Running multiple AI agents together only pays off under specific conditions, according to a new arXiv study on multi-agent task structure. It finds collaboration wins on long-horizon tasks where the subtasks have sparse dependencies on each other, and loses on tightly coupled sequential work.

That matters for anyone deciding whether to build a workflow as one agent or a crew of them. The instinct to default to multi-agent, on the assumption that more agents means more capability, doesn't hold up here.

On sequential work, a single-agent harness stays ahead. Coordinating agents costs context, and that cost has no payoff when each step depends on the one before it in a tight chain. The paper's most useful result is a negative one: enlarging the agent pool or adding more recursion did not reliably improve outcomes.

So the variable that decides whether collaboration helps isn't model strength or agent count. It's whether the task's steps can run somewhat independently of each other. A research task with parallel sub-questions fits the sparse-dependency case. A task where step three needs the exact output of step two, which needs step one, doesn't, and adding agents to it just adds coordination tax on top of a chain that has to run in order anyway.

The study doesn't say how it measured "sparse" versus "tight" dependency across its task set, so anyone applying this to a specific workflow has to make that call themselves. Before scaling an agent fleet, check whether the task actually decomposes into pieces that don't wait on each other, or whether it just looks that way on a whiteboard.