A Multiplicative Handoff Decay Model and Circuit-Breaker Framework for Mitigating Cascading Hallucinations in Multi-Agent Systems

Sequential task delegation in multi-agent LLM systems creates an inherent vulnerability to compounding informational distortion—a phenomenon known as cascading hallucination or the "broken whisper" effect. Minor semantic drifts and ambiguities introduce noise at each agent handoff, culminating in downstream execution failures. This paper introduces a formal mathematical framework to measure and penalize this degradation. By modeling individual agent misinterpretations as a discrete error coefficient we establish a network fidelity metric, that utilizes a sequential product series to quadratically penalize multi-layer degradation. Rather than relying on rigid conservation models, the framework deploys a dual-bound threshold circuit breaker. When message integrity falls outside a defined tolerance interval relative to the original user input , the system automatically halts execution and restarts the communication chain. This replaces reactive prompting with active, deterministic mathematical guardrails to ensure multi-agent workflow reliability.