Researchers have developed ATLAS, a multi-agent framework using large language models for autonomous trading.

Researchers have developed ATLAS, a multi-agent framework using large language models for autonomous trading. The system integrates market data, news, and corporate fundamentals to make trading decisions while operating in an executable action space that produces real market orders. A key innovation is Adaptive-OPRO, a prompt-optimization technique that dynamically adjusts instructions based on real-time market feedback, enabling the system to improve performance over time despite late and noisy reward signals. Testing across multiple equity regimes and LLM families shows that adaptive prompts consistently outperform static approaches, while traditional reflection-based feedback methods failed to deliver systematic improvements. This advancement addresses critical challenges in deploying AI for financial decision-making.