Resilient Consensus in Agentic AI

The researchers model the process by which large language model agents reach agreement as a Byzantine consensus game, conducting controlled experiments on complete graphs and general communication graphs to test whether classical resilient consensus theory, developed for deterministic agents, transfers to LLM agents that may behave adversarially. Results show that prompt-driven LLM agents alone cannot achieve consensus that is achievable in principle; consensus can still fail even in settings where classical theory guarantees the existence of a convergent algorithm, and this failure persists across different temperature and time-horizon settings; wrapping agents with classical resilient consensus filters can improve consistency, with gains depending on the degree of robustness the underlying topology already provides. The authors accordingly argue that classical resilient consensus theory can serve as a useful lens for examining agentic AI safety.