What it does This Python project wires up 19 LLM-powered agents—13 named after real-world investing legends (Buffett, Graham, Taleb, Cathie Wood, etc.) plus valuation, sentiment, fundamentals, technicals, risk, and portfolio managers. You feed it stock tickers, the agents debate, and the system spits out a consensus decision. It runs via CLI with Poetry or through a newer web UI, and includes a backtester for historical simulation.

The interesting bit The gimmick is the gimmick: each agent is literally prompted to channel a specific investor’s philosophy. The “Michael Burry” agent hunts deep value; the “Nassim Taleb” agent obsesses over tail risk. Whether this produces better signals than a single well-prompted model is left as an exercise for the reader, but it’s a clever teaching device for how multi-agent systems can model conflicting viewpoints.

Key highlights

• Supports OpenAI, Anthropic, Groq, DeepSeek, or local Ollama models
• CLI backtester with date-range flags for historical simulation
• New web application interface (separate install under /app)
• Requires FINANCIAL_DATASETS_API_KEY for market data plus at least one LLM key
• MIT licensed, actively seeking small, focused PRs

Caveats

• The system does not execute trades; it’s simulation-only
• README is vague on how agent conflicts actually resolve—portfolio manager gets final say, but the mechanism isn’t detailed
• Web app instructions are offloaded to a subdirectory with no setup detail in the main README

Verdict Worth a spin if you’re teaching or learning multi-agent LLM patterns, or if you want to backtest how a “Buffett vs. Taleb” debate might have performed in Q1 2024. Skip it if you need a real trading bot or rigorous quantitative validation.