Practice: The bull and the bear
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. Why does the system stage a debate instead of just asking one agent whether the company is a good investment?
Show answer
Because a single pass gives you one slanted view with matching confidence; it underweights whatever does not fit its lean. That is what one pass does, not a flaw you can prompt away. Splitting the question into two opposite mandates forces the strongest case on each side into the open, so nothing important gets glossed over.
2. The bull and bear researchers get the same evidence. What is different about them, and what do they both do beyond stating their own case?
Show answer
The difference is the mandate, not the data: one is told to build the case for, the other the case against. Beyond stating its own case, each reads the other’s most recent argument and is told to rebut it. That rebuttal is what makes it a real debate rather than two monologues written in separate rooms.
3. What ends the debate? Describe the check that runs after each turn.
Show answer
A short function compares a turn counter against a round limit. If enough turns have happened, it hands the transcript to the judge (the Research Manager). Otherwise it alternates: if the bull just spoke, the bear is next, and vice versa. By default the limit is one round, so one bull turn and one bear turn, then the judge.
4. Who makes the decision, and why is it important that this agent did not argue either side?
Show answer
The Research Manager, a separate agent, makes the decision. It matters that it did not argue a side because that separates advocacy from authority: a neutral judge reading both cases beats either advocate grading its own work. If the deciding agent were also an advocate, you would just be letting the winner score the match.
5. The judge runs on a different model from the two researchers. Which one, and how does that connect to lesson 1?
Show answer
The judge runs on the more capable (and more expensive) model, while the two advocates run on the cheaper one. That is lesson 1’s idea made concrete: you do not need your best model everywhere, only at the judgment point. Spend capability where the decision is actually made.
6. The judge cannot just say “it depends.” What is it required to produce?
Show answer
It has to commit to exactly one of five stances (Buy, Overweight, Hold, Underweight, Sell), and it is told to reserve the balanced middle stance for when the evidence really is even. The committed decision is written into shared state as the investment plan, the one thing the debate leaves behind for the next agent to act on.
Try it yourself: stage a debate between your own agents
Section titled “Try it yourself: stage a debate between your own agents”No tooling, no cost; this is design judgment. Pick a real decision you could imagine handing to AI (for example: “should our small team adopt this new software tool?” or “is this neighborhood a good fit for a first apartment?”). Then design the debate:
1. Write the two MANDATES. One agent argues FOR, one argues AGAINST. Give each the same evidence, and require each to rebut the other's strongest point.2. Set the TURN CAP. How many turns each before you stop? Why that number?3. Define the JUDGE. A separate agent (your most capable model) reads both cases. List the fixed set of decisions it must choose from (not free text).Show answer (worked example: adopt a new software tool?)
- Mandates: the “for” agent argues the tool saves time, fits the workflow, and is worth the switching cost; the “against” agent argues the migration risk, the learning curve, and the lock-in. Both read the same notes on the team’s current setup, and each must answer the other’s strongest objection.
- Turn cap: one round each is plenty for a reversible call; bump to two rounds for an expensive, hard-to-undo one. Either way the debate ends at the cap, so it cannot spiral.
- Judge: a separate agent that must pick one of: Adopt now, Pilot with one team first, Revisit in six months, or Decline. A fixed menu forces a real decision instead of a hedge.
Notice the shape matches the lesson: opposite mandates over shared evidence, a hard cap on the back-and-forth, and a separate judge that must commit. 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 stage a debate instead of asking one agent for the answer?
A single pass gives one slanted view with matching confidence and underweights whatever does not fit its lean. Splitting the question into opposite mandates forces the strongest case on each side into the open. The fix is structural, not a better prompt.
Q. What is the difference between the bull and bear researchers?
Only the mandate. The bull is told to build the case for; the bear, the case against. They read the same evidence (the four analyst reports). The assignment differs, not the data.
Q. What makes it a debate rather than two monologues?
Each researcher reads the other’s most recent argument and is told to rebut it. So the bull must answer the bear’s strongest point and the bear must answer the bull’s, instead of each writing in isolation.
Q. What ends the bull/bear debate?
A check after each turn compares a turn counter against a round limit. Once enough turns have happened, the transcript goes to the judge. By default it is one round: one bull turn, one bear turn, then the judge.
Q. Who decides, and why does it matter that they did not argue?
The Research Manager, a separate agent. It matters because it separates advocacy from authority: a neutral judge reading both cases beats an advocate grading its own work.
Q. Which model does the judge run on, and why?
The more capable (more expensive) model, while the advocates run on the cheaper one. You spend your best model at the judgment point, not on every voice. That is lesson 1’s deep-versus-fast split made concrete.
Q. What is the judge required to produce?
A commitment to exactly one of five stances (Buy, Overweight, Hold, Underweight, Sell), with the balanced middle reserved for genuinely even evidence. It is saved as the investment plan in shared state.
Q. What is the transferable debate-then-judge pattern?
Assign opposite mandates over the same evidence, cap the number of turns, then hand both cases to a separate judge (ideally your most capable model) that must commit to one option from a fixed set.