Adding more help to a human-AI team can make things worse.
Researchers tested shared-workspace human-AI teams across 1,482 sessions using the Collaborative Gym environment and DiscoveryBench tasks. The setup: AI agents and human collaborators had to coordinate before submitting a final answer. The finding that cuts against the "more is better" assumption — adding relevant collaborators actually lowered performance when teams had no structure for routing who does what. The culprit is process loss, the coordination overhead that eats into any gains from extra capability.
The fix, at least in simulation, was scaffolding: shared group memory combined with human-in-the-loop gates that require a designated participant to approve selected actions. That structure raised mean performance, most noticeably in three-person teams, and pushed expertise toward the decisions where it actually mattered. The implication is that team composition is only half the problem — the other half is how contributions get integrated.
This lands at an awkward moment for AI deployment strategies that lean on adding more agents as a default performance lever. The study did not test real humans, only simulated collaborators, which limits how far the results travel — but the core dynamic, that coordination overhead can cancel out capability gains, is a known problem in human team research too.