Resilient Consensus in Agentic AI

The study explores the application of classical resilient consensus theory to large language model (LLM) agents in multi-agent systems, framing their agreement process as a Byzantine consensus game. Experiments reveal that prompted LLM agents often fail to achieve consensus, even in theoretically favorable conditions, while employing classical consensus filters enhances agreement, contingent on the robustness of the underlying communication topology. This research underscores the importance of integrating classical consensus mechanisms to improve the reliability of agentic AI systems in adversarial environments.