The capstone: cheatsheet

For education only. Multi-agent architecture taught via stock analysis; not investment, financial, or trading advice.

The one idea

The capstone runs the whole team you studied, end to end, in a safe simulation you can run yourself. (The open-source TradingAgents framework, frozen snapshot 7e9e7b8; the same code you read across the track.)

The pipeline at a glance

analysts (x4)  ->  Bull / Bear  ->  Research Manager  ->  Trader
   ->  Aggressive / Conservative / Neutral  ->  Portfolio Manager  ->  done

The two managers run on the deep model; everyone else runs on the faster one.

The seven reports it writes

Report	Who writes it	Lesson
four analyst reports	the tool-using analysts	2
investment plan	the research manager (the judge)	3
trader’s plan	the trader	4
final trade decision	the portfolio manager (the gate)	5

The final decision is one of five stances (Buy, Overweight, Hold, Underweight, Sell): the system’s own label for a written research opinion, not advice, and no orders are placed.

Run it yourself (optional, a simulation)

It writes a research opinion and reports; it places no trades. Use your own AI provider key.

python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .
export ANTHROPIC_API_KEY=your-own-key-here
tradingagents

Then answer the interactive questions (company, date, analysts, depth, provider, model). Reports land in a folder under your home directory.

First-run gotchas

• Python 3.10 or newer; build the environment with 3.11 or 3.12 explicitly.
• Pick Anthropic (Claude) to match the key shown above; if a run errors on the first agent, switch to a Claude model that accepts the effort setting (Claude Sonnet 4.6 works at this snapshot).
• Set a spending cap on your provider dashboard first; a run makes many model calls and costs real money. Check the dashboard for the actual figure.
• A shallow run takes roughly ten to fifteen minutes.
• Never put your provider key in a file you commit or share.

The honest caveat

The capstone runs the real open-source library you studied. The Trading Agents Lab desktop app is a separate, polished product on a simplified engine (a cost and user-experience choice), not the teaching artifact. Both are education and research only.

The transferable pattern

Specialists over one generalist, tool-using data gathering, forced disagreement, deciding separated from doing, a gate with veto power, one shared source of truth, learning from real outcomes, and the capable model spent only at the points of decision.