Your Trading Agent Can't Actually Trade, and Your
TraderBench: AI Trading Agents Use Fixed Strategies Regardless of Market Conditions
Xiaochuang Yuan (Amazon), Hui Xu (Stony Brook), Silvia Xu (Stanford), Cui Zou (University of Oklahoma), and Jing Xiong (UC Santa Cruz) introduce TraderBench, a benchmark evaluating AI agents across knowledge retrieval, analytical reasoning, options trading, and crypto trading — using a two-agent architecture built on the A2A protocol with six MCP servers for financial data access. Submitted to the Agents in the Wild Workshop at ICLR 2026.
The headline finding: 8 of 13 models scored approximately 33 on crypto trading with less than 1 point of variation across four progressive adversarial conditions — from baseline to noise injection to meta-adversarial manipulation. The models were not adapting to market conditions. They were executing fixed strategies regardless of what the market was doing. Extended thinking capabilities improved knowledge retrieval by 26 points but had zero measurable impact on trading performance (+0.3 crypto, -0.1 options). Paper
If your agent needs to make dynamic decisions in adversarial environments, current models are not doing what you think they are doing. They are pattern-matching to fixed strategies, not reasoning about changing conditions.