agents/investor-personas/ — Famous Investor AI Personas

📚 What is this & why does it exist?

These 13 agents are the creative heart of the system. Each encodes decades of a real investor's published philosophy, books, and interview quotes into an LLM system prompt. The agent then uses this persona to interpret financial data — the same P/E ratio that excites a growth investor like Cathie Wood might horrify a value investor like Ben Graham. This diversity of perspective is the core insight: no single philosophy works in all markets.

🤔 Why Is It Designed This Way?

The persona-agent pattern encodes domain expertise as natural-language system prompts rather than hard-coded rules. Each agent follows an identical structural contract — (state: AgentState) → {signal, confidence, reasoning} — but swaps in a different philosophical framework via its prompt. This design was chosen over a single agent with a "style" parameter because each investor's logic requires different sub-analyses: Buffett needs moat analysis while Taleb needs tail-risk metrics. Separating agents also allows LangGraph to run them in parallel across the fan-out phase, maximizing throughput. The trade-off is code duplication across agents, but the domain separation makes each agent independently testable and modifiable.

warren_buffett_agent — Core Analysis Pipeline

def warren_buffett_agent(state: AgentState, agent_id: str = "warren_buffett_agent"):
    """Analyzes stocks using Buffett's principles and LLM reasoning."""
    data = state["data"]
    end_date = data["end_date"]
    tickers = data["tickers"]
    api_key = get_api_key_from_state(state, "FINANCIAL_DATASETS_API_KEY")
    analysis_data = {}
    buffett_analysis = {}

    for ticker in tickers:
        # Step 1: Fetch raw financial data
        progress.update_status(agent_id, ticker, "Fetching financial metrics")
        metrics = get_financial_metrics(ticker, end_date, period="ttm", limit=10, api_key=api_key)

        # Step 2: Run Buffett-specific sub-analyses
        fundamental_analysis = analyze_fundamentals(metrics)
        consistency_analysis = analyze_consistency(financial_line_items)
        moat_analysis = analyze_moat(metrics)
        pricing_power_analysis = analyze_pricing_power(financial_line_items, metrics)
        book_value_analysis = analyze_book_value_growth(financial_line_items)
        mgmt_analysis = analyze_management_quality(financial_line_items)
        intrinsic_value_analysis = calculate_intrinsic_value(financial_line_items)

        # Step 3: Aggregate sub-analysis scores
        total_score = (
            fundamental_analysis["score"] +
            consistency_analysis["score"] +
            moat_analysis["score"] +
            mgmt_analysis["score"] +
            pricing_power_analysis["score"] +
            book_value_analysis["score"]
        )

        # Step 4: Calculate margin of safety (Buffett's key metric)
        margin_of_safety = None
        intrinsic_value = intrinsic_value_analysis["intrinsic_value"]
        if intrinsic_value and market_cap:
            margin_of_safety = (intrinsic_value - market_cap) / market_cap

        # Step 5: Send to LLM with Buffett persona prompt
        buffett_output = generate_buffett_output(
            ticker=ticker, analysis_data=analysis_data[ticker],
            state=state, agent_id=agent_id,
        )

        # Step 6: Store in standard signal format
        buffett_analysis[ticker] = {
            "signal": buffett_output.signal,
            "confidence": buffett_output.confidence,
            "reasoning": buffett_output.reasoning,
        }

    # Write signals to shared state
    state["data"]["analyst_signals"][agent_id] = buffett_analysis