Good Enough is Better: Uncertainty-Aware Aggregation for Multi-LLM Systems

Recent advances in Large Language Models (LLMs) have driven the deployment of complex, multi-stage agentic pipelines. While standard aggregation and multi-agent consensus methods can improve system reliability, sub-agents frequently disagree, and their performance degrades unpredictably under distribution shifts. In this work, we propose the Robust Satisficing Ensemble (RSE), a probabilistic aggregation framework designed to secure multi-agent systems against unpredictable out-of-distribution (OOD) data. By bridging a robust satisficing objective with a quadratic risk surrogate, RSE derives a closed-form analytical inner adversary. This fundamental reformulation enables highly efficient, mathematically guaranteed weight updates without the computational bottleneck of standard minimax optimization. We evaluate RSE across diverse settings, ranging from generative agentic workflows, such as reward modeling under heuristic shifts and selective copying, to standard distribution shift benchmarks (SST-5, TREC, CivilComments, and HuffPost). Our results demonstrate that RSE prevents catastrophic failure and consistently outperforms strong robust baselines, including GroupDRO, IRM, and SRM. These results establish RSE as a practical, theoretically grounded safeguard for reliable agentic routing in dynamic environments.