Does agent confidence actually signal competence in deliberation?

Multi-agent deliberation succeeds or fails not only on individual agents' predictions but on how they communicate and update. Modeling deliberation through Friedkin-Johnsen opinion dynamics — a tractable account of stubbornness, influence, and opinion change — yields a clean reframe: because the FJ parameters are input-dependent, deliberation behaves as a mixture of experts with adaptive routing. That explains when a multi-agent system beats single agents and static ensembles: when routing actually reflects agent competence on the input.

The problem is that competence is latent. In practice influence is established through observable proxies — an agent's self-assessed confidence, its perceived confidence, and its initial alignment with others. None of these is competence. So the routing that gives multi-agent systems their theoretical advantage is driven by the wrong signal, and miscalibrated confidence becomes influence. The paper names the resulting limitations precisely: miscalibrated agent confidence, misleading consensus, and routing errors.

This sharpens the vault's existing multi-agent failure cluster. Since Why do multi-agent LLM systems converge without genuine deliberation?, the FJ/MoE lens supplies the mechanism: agents route influence toward whoever sounds most confident, and consensus forms around persuasion rather than evidence — exactly the pattern When does debate actually improve reasoning accuracy? documents. The design implication is calibration-first: a multi-agent system is only as good as its agents' confidence is honest.