Practice: The risk gate
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. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. The trader already produced a concrete plan. Why does the system add a risk review and a gate instead of stopping there?
Show answer
Because shipping whatever the last step produced is reckless. A careful design stress-tests the plan before it becomes the final answer: it asks “what could go wrong?” through a review, and then routes the result through a gate that can confirm, soften, or overrule it. That review-and-gate step is what separates a careful system from a reckless one.
2. Why three risk seats with different attitudes, rather than one careful reviewer?
Show answer
Because one reviewer (or three who think alike) only catches the concerns that fit their own bias. Three seats with genuinely different attitudes (bold, cautious, balanced), each prompted to challenge the other two, surface failure modes from different directions. The disagreement is the feature: the plan gets pressure from every angle at once.
3. The three seats rotate. What ends the rotation, and what is the default?
Show answer
A short function checks a counter against a round limit after each turn. Once enough turns have happened, it hands off to the Portfolio Manager (the gate). The default is one round, and because there are three seats the counter is compared against three times the limit, so each seat speaks once and then the gate takes over. Same bounded shape as the bull and bear debate.
4. What makes the Portfolio Manager a gate rather than just another opinion?
Show answer
Two things. First, it is a separate role: it argued none of the three risk positions, so it is not grading its own work, and it can confirm, soften, or overrule the trader’s proposal. Second, it has authority and runs on the deep, more capable model. A reviewer produces opinions; a gate makes the final decision and can say no.
5. The system spends its deep model on exactly two agents. Which two, and what do they have in common?
Show answer
The judge after the bull and bear debate (the Research Manager) and the final gate after the risk review (the Portfolio Manager). Both are the moments where the system commits to a decision, and both have authority the agents around them do not. Capability and authority are concentrated at the two decision points, not spread evenly across the team.
6. What can the gate do to the trader’s proposal, and where does its decision go?
Show answer
It can confirm it, water it down, or overrule it entirely. Its decision is one of five stances and is written into shared state as the final trade decision, the last word that the whole pipeline was building toward.
Try it yourself: add a review-and-gate stage to your own system
Section titled “Try it yourself: add a review-and-gate stage to your own system”No tooling, no cost; this is design judgment. Pick a workflow that ends in a real recommendation (for example: an AI that drafts a job offer before it goes out, or one that proposes a marketing budget). Then design the end of it:
1. Design the REVIEW. Name two or three reviewers with genuinely different priorities. What does each one care about most? How do they respond to each other rather than review in isolation?2. Bound the review. How many rounds before it stops?3. Design the GATE. A separate agent (not a reviewer) reads the whole review and issues the final call. What can it do: confirm, modify, reject? Which model does it run on, and why?Show answer (worked example: job offer before it goes out)
- Review: a “candidate advocate” (push to win the hire, generous offer), a “budget guardian” (protect the comp band, watch precedent), and a “fairness reviewer” (consistency with peers). Each responds to the others, so the generous offer is tested against budget and fairness at once.
- Bound: one round each is enough for a routine offer; add a round for a senior or unusual hire.
- Gate: a separate hiring manager agent reads the full review and decides: send as drafted, adjust the numbers, or send back for rework. It runs on your most capable model, because this is where the decision actually commits, and it can overrule any single reviewer.
Notice the shape matches the lesson: a review built on disagreement, bounded, then a separate gate with authority and the best model. The domain changed; the pattern did not.
Flashcards
Section titled “Flashcards”Eight cards. 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. Why add a risk review and gate after the trader's plan?
Because shipping whatever the last step produced is reckless. A review stress-tests the plan (“what could go wrong?”), and a gate decides what to do about the risks, with the power to confirm, soften, or overrule.
Q. Why three risk seats with different attitudes?
One reviewer, or three who think alike, only catches concerns that fit a shared bias. Three different attitudes (bold, cautious, balanced), each challenging the others, surface failure modes from every direction. The disagreement is the feature.
Q. What ends the three-seat rotation?
A counter checked against a round limit after each turn. By default one round: each of the three seats speaks once, then the gate takes over. The factor is three because there are three seats.
Q. What makes the Portfolio Manager a gate, not just an opinion?
It is a separate role that argued none of the three positions, and it has authority: it issues the final decision and can confirm, soften, or overrule the trader’s proposal. A reviewer opines; a gate decides and can say no.
Q. Which two agents get the deep model, and why those two?
The judge after the bull and bear debate and the final gate after the risk review. Both are where the system commits to a decision and both hold authority the surrounding agents do not. Capability and authority sit at the two decision points.
Q. What can the gate do to the trader's proposal?
Confirm it, water it down, or overrule it entirely. Its verdict (one of five stances) is written to shared state as the final trade decision, the last word of the pipeline.
Q. What is the review-and-gate pattern?
Review the work with reviewers who genuinely disagree, bound that review, then put a separate gate at the end with authority to confirm, change, or reject, running on your most capable model.
Q. How do you spot a workflow with no gate?
Nothing reviewed the work with an adversarial eye and nothing had the authority to stop it. That output is a first draft, not a decision.