TradingAgents presents a groundbreaking multi-agent framework for financial trading that leverages large language models (LLMs) to simulate the collaborative dynamics of real-world trading firms. Unlike previous approaches that focused on single-agent systems or independent multi-agent data gathering, this framework organizes specialized LLM-powered agents into roles mirroring actual trading organizations, including fundamental analysts, sentiment analysts, technical analysts, and traders with varying risk profiles. The system incorporates Bull and Bear researcher agents to assess market conditions, a dedicated risk management team to monitor exposure, and traders who synthesize insights from agent debates and historical data to make informed trading decisions.

The framework's architecture enables dynamic collaboration among agents, allowing them to debate market perspectives and collectively arrive at trading decisions. Through extensive experiments, TradingAgents demonstrates superior performance over baseline models, achieving notable improvements in key financial metrics including cumulative returns, Sharpe ratio, and maximum drawdown. These results highlight the potential of multi-agent LLM frameworks to enhance trading performance by better replicating the complex, collaborative nature of professional trading environments. The open-source availability of TradingAgents provides a valuable resource for further research and development in automated financial trading systems.