TradingAgents is an open-source multi-agent LLM framework from Tauric Research that simulates a real-world trading firm using specialized AI agents that collaborate to analyze markets and make trading decisions. The project has rapidly grown to over 74,400 GitHub stars and 14,500 forks, establishing itself as one of the most influential autonomous agent frameworks of 2026. Instead of relying on a single monolithic model, TradingAgents decomposes the trading workflow into distinct professional roles, mirroring how actual investment firms operate. ## Why TradingAgents Matters Most LLM-based trading experiments treat the model as a black box, asking it to output buy or sell signals without structured deliberation. TradingAgents takes a fundamentally different approach by recreating the social and analytical structures of an institutional trading desk inside an agent system. Analyst agents specialize in fundamentals, sentiment, news, and technical signals. Researcher agents argue the bullish and bearish cases. Trader, risk, and portfolio manager agents make the final decisions through structured debate and approval workflows. This division of labor produces more transparent, auditable reasoning than a single-agent setup. The framework is built on LangGraph, giving developers explicit control over agent state transitions, message passing, and debate rounds. The architecture is modular, so individual agent roles can be swapped, fine-tuned, or replaced with custom logic without touching the rest of the pipeline. ## The Analyst Team The analyst team consists of four specialists. The Fundamentals Analyst evaluates company financials, valuation ratios, and intrinsic value estimates. The Sentiment Analyst aggregates news flow and social media discourse to gauge market mood. The News Analyst tracks macroeconomic indicators and event-driven catalysts. The Technical Analyst applies classical indicators such as MACD, RSI, and Bollinger Bands to identify entry and exit points. Each analyst writes a structured report that downstream agents consume. ## Researcher Debate and Risk Management A pair of bullish and bearish researcher agents critically examine the analyst reports through configurable debate rounds. This adversarial process surfaces weaknesses in each thesis before any trade is placed. The Trader agent then synthesizes the debate output into a concrete proposal with position sizing and timing. A Risk Management team evaluates portfolio volatility, liquidity constraints, and exposure limits. Finally, a Portfolio Manager agent approves or rejects the transaction, providing a clear audit trail. ## Multi-Provider LLM Support TradingAgents supports a wide range of LLM providers including OpenAI, Google Gemini, Anthropic Claude, DeepSeek, Qwen, GLM, MiniMax, and xAI Grok. It also integrates enterprise providers like Azure OpenAI and AWS Bedrock, and local models through Ollama. This flexibility lets teams optimize cost, latency, and capability per agent role, for example using a smaller model for the analyst stage and a stronger reasoning model for the portfolio manager. ## Operational Features The framework ships with persistent decision logging that tracks realized returns and alpha versus SPY, checkpoint-based resumption for long backtests, Docker deployment, an interactive CLI with real-time progress tracking, and proxy and remote Ollama compatibility. Backtests preserve date fidelity, preventing look-ahead bias when replaying historical periods. ## Limitations The project explicitly warns that it is not financial, investment, or trading advice and is intended for research and education. Running the full agent pipeline with frontier models can be expensive per decision, especially with many debate rounds. The framework focuses on equities and does not natively handle options, futures, or crypto microstructure. Live execution requires integrating an external broker API, which is not bundled.