Autonomous Agents Achieve 99.9% Trading Success Rate

This research examines the reliability of autonomous language-model agents used in a trading context, specifically within the DX Terminal Pro framework. Over a 21-day deployment, these user-funded agents conducted trades in a bounded onchain market, resulting in over 7.5 million agent invocations and approximately $20 million in trading volume. The agents were tasked with making independent buy/sell decisions under guided constraints provided by users, achieving an impressive 99.9% settlement success rate for policy-valid transactions. The study highlights the importance of an operating layer that includes prompt compilation and policy validations which significantly enhance agent performance beyond just the base model.

Strategically, this research underscores the potential for capital-managing agents to operate with high reliability when the design accommodates comprehensive validation processes. The results reveal a substantial improvement in trading efficiency, from a success rate of fabricated sell rules dropping from 57% to 3%, and a notable increase in capital deployment from 42.9% to 78%. These findings indicate that enhancing the backend operational controls not only improves autonomy but may reduce reliance on manual trading interventions, fostering a more effective digital asset management strategy.