TradingAgents, an open-source multi-agent framework for LLM-driven financial trading, has matured rapidly through nine releases since its initial launch earlier this year. The latest version, v0.4.0, shipped in August 2026 with corrections to look-ahead bias in FRED macroeconomic data, social sentiment feeds, and the decision-log memory, alongside support for GPT-5.6 and GLM-5.3 models. The project, authored by Yijia Xiao, Edward Sun, Di Luo, and Wei Wang, is now a fully open-source framework for researchers studying how specialized AI agents collaborate to evaluate markets.
How the agent hierarchy mirrors a real trading desk
The framework decomposes trading analysis into a chain of specialized LLM agents, each handling a distinct function. A Fundamentals Analyst evaluates company financials and performance metrics to surface intrinsic values and red flags. A Sentiment Analyst pulls from news headlines, StockTwits, and Reddit to produce a single read on short-term market mood. A News Analyst monitors global events and macroeconomic indicators for their market impact. A Technical Analyst applies indicators such as MACD and RSI to detect price patterns and forecast movements.
Above the analysts sit two opposing researchers, one bullish and one bearish, who debate the findings through structured argumentation to balance potential gains against risk. A Trader then synthesizes the analysts' reports and the researchers' debates into concrete trade proposals specifying timing and size. A Risk Management team continuously evaluates portfolio volatility, liquidity, and other risk factors before passing assessment reports to the Portfolio Manager, who ultimately approves or rejects each transaction. Approved orders are routed to a simulated exchange for execution.
Provider support and market coverage
TradingAgents supports a wide range of LLM backends. Providers include OpenAI, Google, Anthropic, xAI, DeepSeek, Alibaba DashScope (international and China endpoints), GLM via Zhipu, MiniMax (global and China), OpenRouter, Ollama for local models, and Azure OpenAI for enterprise deployments. Any OpenAI-compatible server such as vLLM, LM Studio, or llama.cpp can be used through the openai_compatible provider. The framework also supports AWS Bedrock with API-key authentication.
The framework covers any market available through Yahoo Finance, using exchange-suffixed tickers. US stocks like AAPL and SPY, Hong Kong tickers such as 0700.HK, Tokyo's 7203.T, London's AZN.L, Indian markets like RELIANCE.NS, Canadian and Australian listings, Chinese A-shares on Shanghai and Shenzhen exchanges, and crypto pairs such as BTC-USD and ETH-USD are all supported. Company identity and the alpha benchmark resolve automatically per market.
State persistence, checkpointing, and reproducibility
TradingAgents persists two types of state across runs. The decision log, always active, appends each completed analysis to ~/.tradingagents/memory/trading_memory.md. On subsequent runs for the same ticker, the framework fetches realised returns, generates a reflection, and injects past decisions and cross-ticker lessons into the Portfolio Manager's prompt. Checkpoint resume is opt-in via the --checkpoint flag and uses LangGraph to save state after each node, allowing interrupted runs to resume from the last successful step rather than restarting. Per-ticker SQLite databases are stored at ~/.tradingagents/cache/checkpoints/.
Because the system is LLM-driven, two runs of the same ticker and date can produce different results. The variation stems from non-deterministic model sampling and from live data sources that return different content over time. The framework addresses this by resolving company identity deterministically from the ticker before any agent runs, and by grounding price and indicator claims in a verified data snapshot. Lowering the sampling temperature can reduce variance for models that honor it, though the current reasoning-first models largely ignore temperature settings.
Running the framework
Installation is straightforward: clone the repository, create a conda environment with Python 3.12, and run pip install . Alternatively, Docker images are available for standard and Ollama-backed configurations. After setting API keys in a .env file, the interactive CLI is launched with the tradingagents command, presenting a screen to select tickers, analysis date, LLM provider, and research depth. For programmatic use, the TradingAgentsGraph class accepts a configuration dict and returns a trading decision through its propagate() method. Configuration options include the choice of deep and quick thinking models, the number of debate rounds, and the temperature setting.
The project carries a research disclaimer: trading performance varies based on model choice, temperature, data quality, and other non-deterministic factors, and the framework is not intended as financial or investment advice. Backtest results should not be treated as replicable strategy returns but rather as a research scaffold for studying multi-agent analysis in financial markets. Contributions are welcome through bug fixes, documentation, and feature proposals, with past work credited per release in the project's CHANGELOG.