Can AI Agents Run Blockchain Consensus? ETH Zurich Test: Success Rate Only 41.6%

Reaching consensus under the condition that some participants may act maliciously is a core challenge faced by all decentralized systems, known as the Byzantine Fault Tolerance problem in distributed computing. The various consensus mechanisms in blockchain essentially address different variants of this issue. A research team from ETH Zurich directly tested the Byzantine consensus capability of LLM Agents in their paper "Can AI Agents Agree?": multiple Agents repeatedly broadcast proposals and vote through a synchronous fully connected network, with some Agents acting as malicious Byzantine nodes deliberately causing disruption. The team used Qwen3-8B and Qwen3-14B, running hundreds of simulations with different group sizes (4, 8, 16 Agents) and varying proportions of malicious nodes.

Even with no malicious nodes at all, the effective consensus rate was only 41.6% (Qwen3-14B achieved 67.4%, while Qwen3-8B only 15.8%). The more nodes there are, the harder it is to reach agreement, with the success rate dropping from 46.6% with 4 Agents to 33.3% with 16 Agents. After introducing malicious nodes, consensus deteriorated further, with failures mainly manifesting as timeouts and convergence stagnation (loss of liveness), rather than value tampering. Simply mentioning the "possible presence of malicious nodes" in the prompt caused the success rate of Qwen3-14B to drop from 75.4% to 59.1%, even when no malicious nodes were actually present. The paper concludes that reliable consensus is not yet an emergent capability that current LLM Agents can be depended upon for, and a cautious attitude should be maintained toward decentralized deployments that rely on robust coordination.