TradingAgents splits a trading decision across LLM agents that play the roles of a trading firm. An analyst team (fundamentals, sentiment, news, and technical analysts) works in parallel, each with its own tools, then bullish and bearish researchers debate the analysts' reports. A trader agent turns the debate into a trade with timing and size, a risk management team assesses volatility and liquidity, and a portfolio manager approves or rejects the proposal before it goes to a simulated exchange.
The framework is built with LangGraph and runs as a Python package or an interactive CLI that streams each agent's progress. Two model tiers split the work: a quick model serves the analysts, researchers, debaters, and trader, and a deep model serves the research and portfolio managers. Each tier can run on its own provider, for example the managers on Claude and the rest on OpenAI models.
Dated runs are point-in-time: US company statements come from SEC EDGAR as filed on the analysis date, and backtests see only data published by each analysis date. A memory log records every decision, and later runs settle past decisions against the realized return and a regional benchmark, then feed a short reflection back to the portfolio manager. The project describes itself as a research scaffold for studying multi-agent analysis, not as financial advice.
Features
- Analyst team: fundamentals, sentiment, news, and technical (MACD, RSI) analysts run in parallel, each on its own tools
- Researcher debate: bullish and bearish researchers argue over the analyst reports for a configurable number of rounds
- Trader and risk review: a trader proposes timing and size, a risk team assesses it, and a portfolio manager approves or rejects
- Two model tiers: separate quick and deep models, each on its own provider and endpoint
- Markets: any market Yahoo Finance covers, including US, Hong Kong, Tokyo, London, India, China A-shares, and crypto pairs
- Point-in-time data: SEC EDGAR fundamentals as filed, plus FRED macro data, Polymarket, Alpha Vantage, StockTwits, and Reddit sources
- Portfolio-aware runs: pass current holdings so the trader, risk analysts, and portfolio manager work against an actual book
- Memory log: past decisions are settled against realized returns and reflected on in later runs
- Backtesting: runs the pipeline over a grid of tickers and dates and scores decisions on realized alpha, resuming interrupted sweeps
- Checkpoint resume: optional LangGraph checkpoints let a crashed run continue from its last successful step
- Reports: saved reports include a single HTML page for reading in a browser, on a phone, or in print
