EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems

EARS (Explanatory Abstention for Reliable Sub-Agent Modeling) is a framework designed for large-scale multi-agent systems (MAS) to enhance the reliability of sub-agent responses by implementing an inter-agent communication protocol for abstention. By using an ensemble of calibrated LLM-as-a-Judge models, EARS generates structured abstention labels and rationales, allowing sub-agents to better detect failure conditions and provide actionable feedback to a coordinator. Evaluated in a production e-commerce assistant, EARS increased the response pass rate from 68.5% to 78.9%, underscoring its effectiveness in improving the reliability of MAS in enterprise settings.