Practice: The capstone
For education only. This practice is about agent architecture; it is not investment, financial, or trading advice, and nothing here is a recommendation to buy or sell anything.
Self-check
Section titled “Self-check”Six short questions that tie the run back to the architecture. Answer each in your head before opening the collapsible.
1. The status board groups the agents into teams. Name the order the work flows through them.
Show answer
The four analysts gather, then the bull and the bear debate, then the research manager rules, then the trader plans, then the three risk seats stress-test, then the portfolio manager makes the final call. Gather, argue, judge, plan, stress-test, decide.
2. The run writes seven reports. Match each to the agent that produced it.
Show answer
Four analyst reports (the tool-using analysts, lesson 2), the investment plan (the research manager, the judge from lesson 3), the trader’s plan (the trader, lesson 4), and the final trade decision (the portfolio manager, the gate from lesson 5). Each file is the output of an agent you studied.
3. Watching the run, how can you tell which two agents got the more capable model?
Show answer
They are the two judgment points: the research manager, who rules on the debate, and the portfolio manager, who makes the final call. Everyone else (analysts, researchers, trader, risk seats) runs on the faster, cheaper model. Capability is spent only where the system commits.
4. The portfolio manager’s final report is described as reconciling the risk arguments, not averaging them. Why does that distinction matter?
Show answer
Averaging would just blend three opinions into a mushy middle. Reconciling means the gate weighs the bold, cautious, and balanced cases against each other and commits to a calibrated decision it can defend. That is the deep-model judge doing real work, not arithmetic.
5. The capstone is a simulation. What does that mean it does and does not do?
Show answer
It runs the real open-source multi-agent pipeline and writes a set of reports plus a final written recommendation (one of five stances). It does not place any orders, connect to any brokerage, or touch real money. It produces a research opinion, not a trade.
6. The track used stock analysis as its example. What is the part you actually take away?
Show answer
The architecture. Focused specialists, tool-using data gathering, forced disagreement, deciding separated from doing, a gate with veto power, one shared source of truth, learning from outcomes, and the capable model spent only at the decision points. The domain was stocks; the pattern is general.
Try it yourself: run the system (optional)
Section titled “Try it yourself: run the system (optional)”This is the one practice in the track that runs real code, and it is entirely optional. It is a simulation: it writes a research opinion and a set of reports, and it places no trades. You need a computer with Python and your own AI provider key.
1. Create a clean Python environment (Python 3.10+, use 3.11 or 3.12).2. Install the project (pip install -e .) from the pinned snapshot 7e9e7b8.3. Set your OWN provider key in your environment (never in a committed file).4. Set a spending cap on your provider dashboard first.5. Run the single command, answer the interactive questions, and watch the status board fill in.6. Open the seven reports it writes and match each one to the agent and the lesson that produced it.Show answer (what to look for when it finishes)
- The status board should fill in team by team, in the pipeline order you studied.
- The reports folder should hold the four analyst reports, the investment plan, the trader’s plan, and the final trade decision.
- Read the final decision last: notice it is one of the five stances and that it reconciles the three risk views rather than averaging them. That is the gate committing.
- If a run errors on the first agent, switch to a model that supports the effort setting, or take the framework’s defaults. A shallow run takes roughly ten to fifteen minutes and costs a modest amount on your own provider account.
The point is not the specific recommendation. It is watching the architecture you read about actually execute, and recognizing every part.
Flashcards
Section titled “Flashcards”Eight cards reviewing the whole track. Click any card to reveal the answer. Use the Print flashcards button to lay out the full set as one card per page for offline review.
Q. What is the pipeline order the capstone runs?
Analysts gather, bull and bear debate, research manager rules, trader plans, three risk seats stress-test, portfolio manager decides. Gather, argue, judge, plan, stress-test, decide.
Q. Which two agents run on the more capable model, and why?
The research manager and the portfolio manager, the two judgment points. Capability is spent where the system commits to a decision, not spread evenly across the team.
Q. What are the seven reports a run writes?
Four analyst reports, the investment plan (research manager), the trader’s plan, and the final trade decision (portfolio manager). One file per stage, each from an agent you studied.
Q. What does it mean that the capstone is a simulation?
It runs the real pipeline and writes a research opinion plus reports. It places no orders, connects to no brokerage, and touches no real money.
Q. Why does the gate reconcile the risk views instead of averaging them?
Averaging blends opinions into a mushy middle. Reconciling weighs the bold, cautious, and balanced cases and commits to a calibrated, defensible decision. That is the deep-model judge working.
Q. What do you actually take away from this track?
The architecture: specialists, tool use, forced disagreement, deciding separated from doing, a gate with veto power, shared state, learning from outcomes, and capability spent at the decision points. The domain was stocks; the pattern is general.
Q. What do you need to run the system yourself?
A computer with Python (3.10+) and your own AI provider key. Market data needs no key. Set a spending cap first, and never commit your key.
Q. What is the difference between the capstone and the desktop app?
The capstone runs the real open-source library you studied. The desktop app is a separate, polished product on a deliberately simplified engine, a cost and user-experience choice. Both are education and research only.