Confidence Laundering in Agent Systems: Why Uncertainty Needs a Latent Carrier

The paper introduces the concept of "confidence laundering" in agent systems, highlighting how uncertainty can be lost during decision handoffs between components, leading to amplified errors. It proposes the use of "latent uncertainty" as a mechanism to preserve and propagate uncertainty through these interfaces, allowing downstream components to better handle fragile upstream decisions. This approach emphasizes the need for uncertainty-preserving designs in agent systems, which could improve robustness and reliability in AI applications.