Enabling Consensus for Agentic Cooperative Decision-Making in Vehicular Networks

Abstract

Recent advances in Artificial Intelligence (AI), especially agentic AI, are pushing distributed autonomous systems beyond predefined rule execution toward joint decision-making among autonomous participants with potentially different models, preferences, constraints, and value assessments. In such environments, coordinated action requires the participants to commit to a single jointly executable outcome, while Byzantine behavior may disrupt consistency in local views and decisions, thereby degrading efficiency and even compromising safety. In this paper, we study this problem through vehicular networks as a representative and safety-critical entry point for agentic consensus. Existing consensus approaches in vehicular collaboration address some Byzantine scenarios, but are largely limited to homogeneous settings with predefined and shared action spaces, models, preferences, and constraints, and often assume binary or reducible-to-binary decisions. To bridge this gap, we investigate Byzantine fault-tolerant decision-making consensus for agentic environments, with vehicular coordination as the primary instantiation. For homogeneous maneuvers involving continuous-valued decisions, we propose (i) a synchronous Byzantine fault-tolerant consensus protocol with veto support, and (ii) a continuous-value consensus protocol for partially synchronous networks. For heterogeneous decision scenarios, we introduce a novel consensus protocol that converges to a single globally acceptable final decision despite heterogeneity in models and constraints, while remaining resilient to Byzantine faults. Analysis and evaluation in vehicular settings show reduced reliance on Minimum Risk Maneuvers (MRMs) and improved cooperation efficiency under varying degrees of agent heterogeneity.