Proceedings of the Annual Meeting of the Cognitive Science Society

Large language models (LLMs) are increasingly deployed in multi-agent teams (MAS), yet recent works have shown that unstructured multi-agent collaboration often degrades performance in noninsight problems. We examined whether previous findings can be generalized to an insight problem solving paradigm and identified process-level signatures that may explain the performance gap. We compared performance between solo versus two-agent ChatGPT-4o teams to solve situation puzzles in a 2 (Agent Configuration) _ 2 (Solving time) experiment (N = 82). Teams were significantly less accurate than solo agents despite asking questions across broad categories. However, question quality was equivalent between conditions. A multilevel mediation analysis controlling for time decomposed the coordination cost into two components: a protocol failure component (47%), in which inter-agent discussion reduced host-directed questioning, and a residual coordination cost (53%) consistent with impaired convergent integration that persisted after controlling for question volume and quality. Doubling the time budget proportionally increased question volume in both conditions but preserved the team to solo ratio, indicating that the cost is structural rather than driven by limited time. These findings extend multi-agent coordination costs to insight problem solving and suggest that unstructured multi-agent interaction supports the divergent phase but disrupts the convergent phase.