In multi-agent systems built from large language models, each agent's decisions depend on information produced by another, so uncertainty about that information must survive every hand-off. Existing uncertainty quantification only measures a model's confidence in its own output and ignores uncertainty originating elsewhere. We isolate this gap as a failure mode we call *vanishing uncertainty*, the attenuation of an uncertainty signal between the agent that produces it and the agent that consumes it. In a controlled two-agent protocol where an orchestrator relays a subagent's reply, we measure four standard estimators at both ends of the hand-off across three open-weight models and tasks spanning parametric knowledge, magnitude estimation, and tool use. Orchestrator and subagent uncertainty correlate weakly even when the relayed content is fixed, showing that standard estimators fail to expose risk whose root lies outside the measured agent. Uncertainty propagation in multi-agent systems is therefore a distinct problem from single-model uncertainty quantification.
@inproceedings{emde2026lost,
title={Lost in Communication: Uncertainty Propagation in Multi-Agent Systems},
author={Cornelius Emde and Anmol Goel and Sangdoo Yun and Seong Joon Oh and Martin Gubri},
booktitle={ICML 2026 Workshop on Statistical Frameworks for Uncertainty in Agentic Systems},
year={2026},
}