AgentRaft: Fault-Tolerant Multi-Agent Consensus Protocol for AI Swarms

Abstract

Multi-agent AI systems suffer from two critical failure modes: Byzantine faults (hallucinations producing incorrect or malicious proposals) and node failures (API timeouts causing silent data loss). AgentRaft applies Raft-inspired distributed consensus principles to AI agent swarms through a 3-level architecture. Level 1 (Protocol Layer) defines an LLM-agnostic, chain-agnostic smart-contract identity standard where agents register keys and stake tokens, a strict JSON message schema (PROPOSAL | VOTE | CHAT | VOTE_NEW_LEADER), and quorum rules (2/3 majority for proposal execution). Level 2 (Orchestration Layer) provides an append-only immutable log via 0G Storage for cryptographic proof of agent decision-making, a state machine that monitors heartbeats and routes VOTE_NEW_LEADER events to a blockchain smart contract, and synchronization of 0G network state back to agents. Level 3 (Application Layer) demonstrates a DeFi Treasury Guardian using LangGraph/AutoGen where a GPT-4o Leader/Proposer agent, a Claude 3 Risk Assessor, and a local-model Compliance agent collaborate; if two follower agents reject the leader proposal, they sign a triggerLeaderElection() transaction on 0G Chain, blocking the DeFi action and recording the censure on-chain. The research question is: can Raft-style consensus mechanisms reliably detect and recover from AI agent Byzantine faults at production latency and cost, and what are the formal correctness bounds?