ETH Zurich Test: AI Agents Achieve Blockchain Consensus Only 41.6% of the Time, Harder With More Nodes

One core challenge for decentralized systems is reaching consensus when some participants act maliciously (Byzantine fault tolerance), which blockchain consensus mechanisms aim to solve. ETH Zurich's team tested LLM Agents' consensus ability using Qwen3-8B and Qwen3-14B models, simulating different node sizes and malicious node ratios.

Results show that even without malicious nodes, the overall consensus success rate is only 41.6% (67.4% for Qwen3-14B, 15.8% for Qwen3-8B). More nodes mean lower success: 46.6% with 4 nodes, dropping to 33.3% with 16 nodes. Mentioning or adding malicious nodes reduces success. The conclusion states current AI Agents' consensus ability is not reliable, so decentralized deployments need caution.