Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in Clinical Multi-Agent Systems

Multi-agent clinical decision systems are susceptible to social conformity, where a holdout agent adopts an incorrect peer consensus in 38% of cases. While Gemini models resist cues in isolation, false assertions and pre-screen flags trigger shortcut cascades across text, imaging, and tabular datasets. Oversight agents fail to distinguish between honest agreement and erroneous adoption, yielding a 100% false-positive rate.