TradingAgents: Multi-Agent AI Framework That Models a Hedge Fund

Most AI trading projects are a single LLM fed data and asked to "make a decision." TradingAgents (GitHub, arXiv: 2412.20138) takes a different path: instead of one agent — an entire trading firm staff, where each role is a separate LLM agent with its own tools, data sources, and prompt.

Built on LangGraph, the framework creates a system that doesn't just "look at data" — it debates with itself before making a decision.

Architecture: From Data to Decision

The full pipeline is a directed acyclic graph (DAG) of 12 nodes:

Analyst Team → Research Debate → Trader → Risk Debate → Portfolio Manager → BUY/HOLD/SELL

Analyst Team: Four Specializations

Research Debate: Bull vs Bear

After analysts produce reports, adversarial debate begins:

1. Bull Researcher builds the bullish thesis
2. Bear Researcher builds the bearish counter-thesis
3. Multi-round debate (configurable via max_debate_rounds)
4. Research Manager (deep thinking LLM) synthesizes both positions

Risk Management: Triple Filter

The trader's proposal passes through three risk managers debating each other:

Portfolio Manager: Final Decision

Receives trader proposal + risk debate results + reflective memory from past decisions. Approves, rejects, or adjusts the trade.

Reflective Memory

The system maintains a decision log. On each subsequent run for the same ticker, it fetches realized returns (raw + alpha vs SPY), generates reflections, and injects history into the Portfolio Manager prompt — creating learning from mistakes without fine-tuning.

Tech Stack

Links

• 💻 GitHub · 📄 arXiv · 🎬 Demo