TradingAgents is a multi-agent financial trading framework designed to simulate the dynamics of real-world trading firms by leveraging specialized Large Language Model (LLM)-powered agents. The framework's core methodology decomposes complex trading tasks into distinct, collaborative roles to achieve a robust and scalable approach to market analysis and decision-making. It is built using LangGraph, ensuring flexibility and modularity, and supports multiple LLM providers including OpenAI, Google, Anthropic, xAI, OpenRouter, and local models via Ollama.

The framework orchestrates a multi-stage decision-making process involving several distinct agent teams:

1. Analyst Team: This team comprises specialized LLM agents responsible for comprehensive market data analysis.
   • Fundamentals Analyst: Evaluates company financial statements and performance metrics to identify intrinsic values and potential risks.
   • Sentiment Analyst: Analyzes social media and public sentiment using sentiment scoring algorithms to gauge short-term market mood.
   • News Analyst: Monitors global news and macroeconomic indicators, interpreting their potential impact on market conditions.
   • Technical Analyst: Utilizes technical indicators, such as Moving Average Convergence Divergence (MACD) and Relative Strength Index (RSI), to detect trading patterns and forecast price movements.

2. Researcher Team: Composed of two specialized agents that critically assess the insights provided by the Analyst Team through structured debates.
   • Bullish Researcher: Focuses on identifying potential gains and positive market catalysts.
   • Bearish Researcher: Focuses on identifying inherent risks and negative market indicators.

This adversarial debate process balances potential gains against inherent risks, refining the market outlook.

3. Trader Agent: This agent synthesizes the comprehensive reports and insights generated by both the Analyst and Researcher teams. Its role is to make informed trading decisions, specifically determining the optimal timing and magnitude of trades based on the aggregated market intelligence.

4. Risk Management and Portfolio Manager: This critical component continuously evaluates portfolio risk by assessing factors such as market volatility and liquidity. The Risk Management team adjusts trading strategies based on ongoing assessments and provides reports to the Portfolio Manager. The Portfolio Manager holds the final authority to approve or reject transaction proposals. If approved, the order is sent to a simulated exchange for execution.

Technically, the framework exposes a Python API through the TradingAgentsGraph object. A trading decision for a specified ticker and date is initiated by calling the propagate() method, e.g., $_ , decision = ta.propagate("NVDA", "2026-01-15")$. The configuration can be extensively customized, allowing users to select llm_provider (e.g., "openai", "google"), assign different LLMs for complex reasoning (deep_think_llm) and quick tasks (quick_think_llm), and set parameters like max_debate_rounds. For instance, $config["deep_think_llm"] = "gpt-5.2"$ and $config["quick_think_llm"] = "gpt-5-mini"$ demonstrate the assignment of models based on cognitive load.

The project is continually updated, with recent releases expanding model coverage to include GPT-5.x, Gemini 3.x, Claude 4.x, and Grok 4.x, alongside architectural improvements for enhanced stability and multi-provider LLM support. While designed for research purposes, the paper explicitly states that trading performance may vary due to factors like chosen backbone models, model temperature, trading periods, and data quality, and it is not intended as financial advice.