Lost in Communication: Uncertainty Propagation in Multi-Agent Systems

Abstract

In multi-agent LLM systems each agent’s decisions depend on information produced by another, yet existing uncertainty quantification only measures a model’s confidence in its own output. We expose a failure mode we call vanishing uncertainty: the uncertainty signal attenuates between the producing agent and the consuming agent. Across three open-weight models and tasks spanning parametric knowledge, magnitude estimation, and tool use, orchestrator and subagent uncertainties correlate weakly even when the relayed content is fixed - showing that uncertainty propagation in multi-agent systems is a distinct problem from single-model UQ.