Bowen Li, Yanchen Qiao, Keyu Lu, Dongcen Ji, Guanlin Li

Large language model (LLM) agents increasingly collaborate through message exchange and tool execution, where unreliable coordination can propagate incorrect actions into external environments. Existing confidence-based coordination methods remain vulnerable because agents control both their proposals and self-reported confidence. We introduce EC-TAC, an evidence-carrying and topology-aware coordination protocol for tool-using LLM multi-agent systems. EC-TAC replaces self-reported confidence with verifiable Evidence Records that bind canonicalized tool actions to schema validation or execution receipts. Each receiver verifies action structure, evidence validity, origin identity, and propagation lineage. Replayed messages cannot increase support, and conflicting records from the same origin are excluded. Decisions are made only when sufficient independent authenticated evidence is observed through the communication topology. We evaluate EC-TAC on selected BFCL and API-Bank tool-use tasks under controlled weak-model replacement scenarios and compare it with CP-WBFT and majority-based coordination. At, EC-TAC attains 91.43% for both ordinary and Strict Robust Accuracy. CP-WBFT attains 59.29% ordinary and 38.57% Strict accuracy. These results show that evidence-based verification and topology-aware propagation improve reliable and auditable coordination for LLM multi-agent systems.

Paper In "AAAI"｜Click Here To Get The Full Paper