Multi-Agent Teams Hold Experts Back

Multi-agent LLM systems are increasingly deployed as autonomous collaborators, where agents interact freely rather than execute fixed, pre-specified workflows. In such settings, effective coordination cannot be fully designed in advance and must instead emerge through interaction. However, most prior work enforces coordination through fixed roles, workflows, or aggregation rules, leaving open the question of how well self-organizing teams perform when coordination is unconstrained. Drawing on organizational psychology, we study whether self-organizing LLM teams achieve strong synergy, where team performance matches or exceeds the best individual member. Across human-inspired and frontier ML benchmarks, we find that---unlike human teams---LLM teams consistently fail to match their expert agent's performance, even when explicitly told who the expert is, incurring performance losses of up to 41.1\% on ML benchmarks. Decomposing this failure, we show that expert leveraging, rather than identification, is the primary bottleneck. Conversational analysis reveals a tendency toward integrative compromise---averaging expert and non-expert views rather than appropriately weighting expertise---which increases with team size and correlates negatively with performance. Interestingly, this consensus-seeking behavior improves robustness to adversarial agents, suggesting a trade-off between alignment and effective expertise utilization. Our findings reveal a significant gap in the ability of self-organizing multi-agent teams to harness the collective expertise of their members.

To tackle hard problems, people increasingly connect several AI models and let them "discuss" before committing to an answer, hoping the team will outperform any single model, much as human teams often outperform individuals. But does teaming actually help? We put groups of AI models through classic group-decision exercises (like ranking survival gear after a crash) and demanding knowledge tests, designing each group so that one member genuinely knew more than the rest. Surprisingly, the teams almost never matched that member — even when we told the group exactly who the expert was — on the hardest tasks losing up to 40% of the accuracy they could have reached by simply trusting whichever member was right. Instead of following the expert, the AIs tended to "split the difference," averaging everyone's opinions, which only worsened as the teams grew larger. The same compromising habit had an upside: it made groups resistant to an adversarial member trying to sabotage the answer. Our findings are a caution for today's rush toward multi-agent AI: adding more models and letting them confer does not guarantee better decisions. In fact, the team can end up holding its expert back.