Consensus-Driven Metacognition in Multi-Agent Systems: A Logic-Based Byzantine Fault-Tolerant Protocol

Abstract

Large language model (LLM) agents fail in a distinctive way: they produce confident wrong answers. Recent work has begun adapting Byzantine fault tolerance (BFT) to multi-LLM networks — notably the weighted BFT protocol of Wang et al. [1] and the Aegean consensus engine [2] — typically by treating agent trust as a scalar that flows continuously through the protocol.

This paper takes the opposite stance: discretization is a feature, not a bug. MBFT (Metacognitive BFT) commits a swarm decision via a small, finite ladder of confidence tiers, defeasible counterproofs, and a reputation-gated veto. The resulting protocol is auditable, deterministically replay able, and its safety / liveness claims are mechanically checkable — properties that continuous Bayesian aggregators struggle to provide. Bayesian and continuous trust schemes remain the right tool for noisy, well calibrated, high-volume regimes (recommender systems, sensor fusion, market making); we argue that high-stakes reasoning swarms — legal, medical, safety-critical agentic deployments — belong to the discrete, defeasible regime that MBFT formalizes. We accompany the paper with an open-source reference implementation whose property suite encodes each theorem as an executable test.