New Study: Multi-Agent AI Teams Act as Mixture-of-Experts

A new paper from Franka Bause, Jonas Niederle, Martin Pawelczyk, and Rebekka Burkholz (arXiv 2605.25929) reveals that multi-agent LLM systems can be understood through the lens of Friedkin-Johnsen (FJ) opinion dynamics—a classic social science model for stubbornness, influence, and opinion change. The authors demonstrate that the FJ parameters depend on the input, meaning each deliberation among agents effectively selects a different mixture-of-experts. Crucially, this explains why multi-agent deliberation often beats both single agents and static ensembles: when the system can route tasks to the most competent agents (the 'influencers'), overall performance improves.

The challenge is that agent competence is latent—you can't directly observe which agent is best for a given input. Instead, the paper analyzes three observable proxies that determine influence: an agent's self-assessed confidence, how confident other agents perceive it to be, and how aligned its initial view is with the majority. By modeling these dynamics, the framework provides a theoretical grounding for why certain agents become 'influencers' in a team. Practitioners can use these insights to design better multi-agent workflows, such as by giving more weight to agents that consistently show calibrated confidence, or by encouraging diversity of opinion to avoid groupthink.