Practice: Why split one AI into many
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 team has twelve agents (with all four analysts on). What principle decides how the work is split?
Show answer
Function. The agents are grouped so each does a distinct kind of work: analysts gather information, researchers argue, a trader synthesizes, risk reviewers stress-test, and managers judge. The team is one workflow broken along its natural steps; the headcount falls out of that, it is not the goal.
2. Name the five groups and what each does.
Show answer
Analysts (gather raw information), researchers (argue opposite sides, a bull and a bear), trader (turn the argument into a concrete plan), risk reviewers (stress-test the plan from three temperaments), and managers (judge: one rules the debate, one makes the final call).
3. The work moves through the team in a clear order. What is it?
Show answer
Analysts gather, then the bull and bear argue, then the research manager rules on the argument, then the trader turns the ruling into a plan, then the three risk reviewers stress-test it, then the portfolio manager makes the final call. The two judges sit at two different points in the flow, not side by side.
4. The system runs on two model tiers. Which roles get the more capable model, and why those?
Show answer
Exactly two: the research manager and the portfolio manager, the two judges. They are the points where one agent weighs everyone else’s work and commits, so a wrong call there is the most expensive. Capability is a budget; you spend it where decisions are made.
5. A teammate says “so the lesson is that you need twelve agents.” What is the more precise takeaway?
Show answer
The number is not the design. Twelve is just the count when all four (optional) analysts are on. The real, steady design is: break the workflow into roles by function, and give your most capable model only to the small number of judgment points. Change the analyst count and the headcount shifts, but the capable model still goes to exactly the two judges.
6. Why does this matter even if you never build a trading system?
Show answer
Because it is how you judge what an AI builds for you. If you know what a sound team looks like (work broken into focused roles, the best model spent only at the decision points), you can look at what an AI hands you and tell whether the structure is right or just busy: ask for a missing reviewer, or question why every step burns your most expensive model. Good structure going in is how you get good results coming out.
Try it yourself: decompose a workflow and place the judges
Section titled “Try it yourself: decompose a workflow and place the judges”No tooling, no cost; this is design judgment. Pick a task you understand well that is not stock analysis (for example: triaging incoming support tickets, or producing a vetted research brief). Then:
1. List the distinct FUNCTIONS the work contains (gather, analyze, argue, synthesize, check, decide). Each becomes a candidate role.2. For each role, write its one-line job and the few tools it would need.3. Circle the JUDGMENT POINTS: where one agent commits on behalf of the rest. Those get your most capable model; the rest run on a cheaper one.Show answer (worked example: support-ticket triage)
- Functions / roles: an intake reader (summarize the ticket), a history gatherer (pull this customer’s prior tickets), a severity classifier, a routing proposer, and a triage judge that commits the final priority and queue.
- Tools: the history gatherer needs ticket-search; the others mostly reason over what the earlier roles produced.
- Judgment point: the triage judge is the one place a wrong call is expensive (a mis-routed urgent ticket). Give it your strongest model; the readers and classifiers can run on a cheaper, faster one.
Notice the shape matches the lesson: many cheap gatherers and analyzers, one or two expensive deciders. 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. What principle decides how a real agent team is split?
Function. The workflow is broken into roles that each do a distinct kind of work (gather, argue, synthesize, stress-test, judge). The agent count falls out of that; it is not the goal.
Q. What are the five groups in this framework?
Analysts (gather information), researchers (a bull and a bear who argue), a trader (synthesize into a plan), risk reviewers (stress-test), and managers (judge: one rules the debate, one makes the final call).
Q. Why a separate bull agent and bear agent instead of one balanced agent?
A single agent asked to weigh pros and cons tends to talk itself into a tidy, balanced answer. Two opponents each argue one side at full strength, so both cases get made properly.
Q. What order does the work flow through the team?
Analysts gather, then bull and bear debate, then the research manager rules, then the trader plans, then three risk reviewers stress-test, then the portfolio manager decides. The two judges sit at different points in the flow.
Q. Which roles get the more capable model, and how many?
Exactly two: the research manager and the portfolio manager, the two judges. Every other role runs on the cheaper, faster model.
Q. What steady rule matters more than the number twelve?
The capable model always goes to exactly the two judges, no matter how many analysts you run. Break the work into roles by function, then spend capability only at the judgment points.
Q. Why does this lesson matter even if you never build a trading system?
It is how you judge what an AI builds for you. Knowing what a sound team looks like lets you tell whether the structure is right or just busy. Good structure in, good results out.
Q. What are the first two steps to design your own agent team?
- Break the workflow into functions, each a role with narrow instructions and few tools. 2. Find the judgment points (where one agent decides for the rest) and put your best model only there.