TradingAgents: Multi-Agent LLM Trading Framework

What TradingAgents Is

TradingAgents is a multi-agent framework that mirrors a trading firm's internal structure: Fundamentals, Sentiment, News, and Technical analysts each read a different data slice, bullish and bearish researchers debate their findings, a Trader agent turns that debate into a decision, and a Risk Management team plus a Portfolio Manager approve or reject it before it reaches a simulated exchange.

How the Multi-Agent System Works

• ✓Analyst Team: Fundamentals Analyst reads company financials, Sentiment Analyst aggregates news headlines, StockTwits, and Reddit chatter, News Analyst tracks macro events, and Technical Analyst runs indicators like MACD and RSI.
• ✓Researcher Team: bullish and bearish researchers run a structured debate over the Analyst Team's output before it reaches a decision.
• ✓Trader Agent composes the analyst reports and researcher debate into a decision on trade timing and size.
• ✓Risk Management team assesses volatility and liquidity, then reports to a Portfolio Manager who approves or rejects the order before it's sent to the simulated exchange.
• ✓Built on LangGraph, with graph-shape-aware checkpoint resume (added in v0.3.1) so a crashed multi-agent run can pick back up instead of restarting.
• ✓A decision log at ~/.tradingagents/memory/trading_memory.md that's always on: each run's realized return (raw and alpha vs SPY) and a one-paragraph reflection get injected into the next run's Portfolio Manager prompt.

Setting Up TradingAgents

Clone the repo (git clone https://github.com/TauricResearch/TradingAgents.git), create a Python 3.12 environment (conda create -n tradingagents python=3.12 && conda activate tradingagents), then run pip install . from inside the project directory. The Docker path skips the local environment: copy .env.example to .env, add your API keys, then docker compose run --rm tradingagents (or docker compose --profile ollama run --rm tradingagents-ollama for local models). Either way you need at least one LLM provider key set — OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, XAI_API_KEY, DEEPSEEK_API_KEY, the DashScope/Zhipu/MiniMax international-or-China key pairs, or OPENROUTER_API_KEY — plus ALPHA_VANTAGE_API_KEY for market data. AWS Bedrock needs the pip install ".[bedrock]" extra plus AWS credentials and a Bedrock model ID; Azure OpenAI needs a copied .env.enterprise; Ollama and any OpenAI-compatible server (vLLM, LM Studio, llama.cpp) need no API key at all.

Running TradingAgents

The interactive path is the CLI: run tradingagents (or python -m cli.main from source) and pick a ticker, analysis date, LLM provider, and research depth from the prompts, then watch the agent team's progress in the terminal. Tickers resolve against whatever market Yahoo Finance covers, using exchange suffixes like 0700.HK, 7203.T, AZN.L, RELIANCE.NS, 600519.SS, or BTC-USD. From Python, import TradingAgentsGraph, copy DEFAULT_CONFIG, and call ta.propagate("NVDA", "2026-01-15") to get a decision back; the same config dict sets llm_provider, deep_think_llm, quick_think_llm, and max_debate_rounds. Add --checkpoint to a CLI run (or checkpoint_enabled: True in config) to persist LangGraph state per ticker in a SQLite database, so a crashed run resumes instead of restarting; --clear-checkpoints resets that state.

Comparing Financial LLM Frameworks

Common Questions

The problem it solves

Testing whether splitting a trading call across specialized roles — a fundamentals read, a sentiment read, a technical read, a bull/bear debate, then a risk sign-off — changes the outcome versus one LLM prompted to just decide, normally means wiring four or five agents and a debate loop yourself. TradingAgents ships that decomposition as running code: the analyst/researcher/trader/risk split from a trading desk, already built on LangGraph.

Best use cases

• •Studying how decomposing a trading call across specialized LLM roles (analyst debate, bull/bear research, risk sign-off) changes the outcome versus a single-prompt approach.
• •Running a decision pipeline against a fixed ticker-and-date pair, e.g. ta.propagate("NVDA", "2026-01-15"), while the price and indicator window stays pinned to that date.
• •Comparing how different LLM providers and models — GPT-5.x, Gemini, Claude, DeepSeek, Qwen, or a local Ollama model — perform inside the same multi-agent pipeline via the config's llm_provider switch.
• •Investigating reproducibility and non-determinism in LLM-driven financial decision systems using the documented temperature setting and checkpoint controls.
• •Running the whole pipeline against a local or self-hosted model (Ollama, vLLM, LM Studio, llama.cpp) instead of a hosted API, for teams that don't want trading data leaving their infrastructure.

Who should try it — and who should skip

Reach for TradingAgents if you're a developer or researcher who wants a working multi-agent pipeline to study how specialized LLM roles debate a trading decision, or a documented multi-provider agent framework to extend. Skip it if you want a production trading bot, live broker execution, or a system with a proven return — the README's disclaimer says this is a research tool, not financial or investment advice, and backtest results aren't guaranteed to match any published figure.

Related repositories