73.7K stars.
That's TradingAgents' number on GitHub. In the AI trading space, it's the most-starred project — well ahead of second place.
But high star count doesn't mean it makes money. After studying this project for a few days, my conclusion is: the architecture is worth learning, but live trading requires caution.
The Architecture
TradingAgents' core idea splits financial analysis into specialized roles:
- Fundamentals Analyst: Analyzes company fundamentals
- Sentiment Analyst: Analyzes market sentiment (news, social media)
- Technicals Analyst: Technical indicator analysis
- Risk Manager: Evaluates risk, controls position size
- Trader: Makes final trading decisions
Each role is an independent agent. They analyze data separately, then aggregate for the Trader Agent to make the final call.
v0.2.5 just dropped (6 hours ago), adding sentiment analyst, env-var config, and more provider support.
Does It Run?
Yes. I got the demo running on an M2 Pro.
Supported models are numerous: OpenAI, Anthropic, DeepSeek, Qwen, GLM, Azure OpenAI, even Ollama local models. You don't necessarily need to pay for APIs — if you have a local GPU, running entirely local with Ollama is possible.
But there's a catch: signal quality heavily depends on model capability. Results from smaller models differ significantly from flagship models. The project doesn't provide clear benchmark data — you need to verify yourself.
What's Worth Learning
Even if you don't trade, TradingAgents' multi-agent design is instructive:
- Role separation: Analysis, risk control, and decision-making are split, avoiding "one brain thinking about everything"
- Information flow design: Each agent focuses on its own data source, reducing noise
- Independent risk layer: The Risk Manager is a dedicated agent whose job is to say "no" — a design missing from many agent systems
If your workflow involves "analyze → decide → execute," this layered architecture is directly借鉴able.
Risk Warning
Must be clear: this is not an "install and make money" tool.
- Past performance ≠ future results: Backtesting well doesn't mean live trading profits
- Model hallucination: LLMs fabricate data and analysis — fatal in financial contexts
- Market risk: No analysis tool predicts black swans
TradingAgents is more of a research framework than a production-grade trading system. Fine as a learning tool, but entrusting real money to it — at least in 2026 — feels too early.
GitHub: github.com/TauricResearch/TradingAgents | 73.7K stars | v0.2.5 (2026-05-12) Based on project code, docs, and actual testing. Not investment advice.