Most attempts to put an LLM in charge of trading hand a single model a price chart and a prompt. TradingAgents takes the opposite bet: that good trading decisions come from disagreement, not a lone oracle. It reconstructs the org chart of a real fund — analysts, a bull-vs-bear research desk, a trader, and a risk committee — and lets the LLMs argue their way to a position, with the debate transcript doubling as the audit trail.
What Sets It Apart
- Decisions emerge from structured conflict: bullish and bearish researchers debate the analysts' findings before anything is traded, so a thesis has to survive a counter-argument rather than a single model's confidence.
- Roles map to real desk functions — fundamentals, sentiment, news and technical analysts, then trader, risk team and portfolio manager — making each step inspectable instead of one opaque "buy/sell" output.
- It is LLM-agnostic to an unusual degree: OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, GLM, Ollama and more, so you can pair a cheap model for the analysts with a stronger one for the final call.
- Memory and checkpointing let a run resume and draw on past decisions instead of starting cold each session.
Who It's For
Great fit if you want a research scaffold for studying multi-agent decision-making, or a transparent base to prototype LLM trading strategies where you can read why a trade was proposed. Look elsewhere if you expect a turnkey profitable bot: the authors are explicit that this is a research framework, not investment advice, and real-world results depend on your models, data feeds and the markets themselves.
