ETH Zurich Practical Test of AI Agent Blockchain Consensus Ability: Success Rate Only 41.6%

Gate News, March 9 — A research team from ETH Zurich tested the Byzantine consensus capabilities of LLM agents in their paper “Can AI Agents Agree?”. The background of the study is that reaching agreement under conditions where some participants may act maliciously is a core challenge faced by all decentralized systems. Various consensus mechanisms in blockchain fundamentally address different variants of Byzantine fault tolerance.

The team used Qwen3-8B and Qwen3-14B models to run hundreds of simulations across different group sizes (4, 8, 16 agents) and proportions of malicious nodes. In the tests, multiple agents repeatedly broadcast proposals and vote through a fully connected synchronous network, with some agents acting as malicious Byzantine nodes to intentionally disrupt.

Results showed that even with no malicious nodes, the effective consensus rate was only 41.6% (67.4% for Qwen3-14B, and just 15.8% for Qwen3-8B). As the number of nodes increased, reaching consensus became more difficult, with success rates dropping from 46.6% at 4 agents to 33.3% at 16 agents. When malicious nodes were introduced, consensus further deteriorated, mainly failing due to timeouts and convergence stalls (loss of activity), rather than data tampering. Simply mentioning “possible malicious nodes” in prompts reduced Qwen3-14B’s success rate from 75.4% to 59.1%, even when no malicious nodes were present.

The paper concludes that reliable consensus is not yet an emergent capability of current LLM agents, and caution should be exercised in decentralized deployments that rely on robust coordination.