Autonomous AI Agents Trade $20M in ETH on Onchain Market

A recent study published on arXiv (2604.26091) investigates the dependability of autonomous language-model agents that convert user requests into verified tool actions in a real capital environment. Conducted over 21 days in DX Terminal Pro, the deployment involved 3,505 user-funded agents trading actual ETH within a limited onchain market. Users set up vaults using structured controls and natural-language methods, while only the agents performed trades. The system recorded 7.5 million agent invocations, approximately 300,000 onchain actions, around $20 million in trading volume, over 5,000 ETH utilized, nearly 70 billion inference tokens, and a 99.9% success rate for policy-compliant transaction settlements. Long-term agents made thousands of sequential decisions, including over 6,000 prompt-state-action cycles, generating a comprehensive trace from user request to settlement. The reliability stemmed from more than just the base model.