Why split one AI into many: cheatsheet

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

The one idea

A real agent team is one workflow broken into specialist roles by function. Spend your most capable model only at the judgment points (where one agent decides for the rest), and run everything else lean. (Anchored to the TradingAgents framework, frozen snapshot 7e9e7b8.)

The roster (twelve agents, by function)

Group	Count	Function
Analysts	4	Gather raw information (market, social, news, fundamentals)
Researchers	2	Argue opposite sides: a bull and a bear
Trader	1	Turn the argument into a concrete plan
Risk reviewers	3	Stress-test the plan: aggressive, conservative, neutral
Managers	2	Judge: the research manager rules the debate, the portfolio manager makes the final call

Twelve is the count with all four analysts on (the default). Analysts are optional; the other eight roles are always present.

The order the work flows

Analysts gather, bull and bear argue, research manager rules, trader plans, three risk reviewers stress-test, portfolio manager decides. The two judges sit at two different points in the flow.

The two model tiers

• Two tiers: a more capable (costlier, slower) model and a more affordable (cheaper, faster) one.
• The capable model goes to exactly two roles: the research manager and the portfolio manager (the judges).
• Every other role runs on the affordable model.
• The steady rule: change the analyst count and the headcount shifts, but the capable model always goes to exactly those two judges. That, not the number twelve, is the design.

How to follow along

The framework is free and open to read at github.com/TauricResearch/TradingAgents (pinned snapshot). No account, no Git, no programming knowledge needed: open the link and read the files in your browser. The short code 7e9e7b8 marks one frozen version, so the lessons always match what you see.

Build your own (the transferable pattern)

1. Break the workflow into functions; make each a role with narrow instructions and few tools.
2. Find the judgment points (where one agent commits for the rest) and put your best model only there. Expect far fewer of them than you would guess.

Pitfalls to dodge

• Counting agents instead of functions (the count falls out of the design; it is not the design).
• Using one model tier everywhere (slow and costly across the board, unreliable if a weak model sits at the judge).
• Assuming more roles means more intelligence (more specialized and more complex, not smarter).

Words to use precisely

• Role: one agent with one focused job (gather, argue, plan, check, decide); the unit a team is built from.
• Judgment point: a place where one agent decides on behalf of the others; where you spend capability.
• Model tier: the capable, costlier model versus the affordable, faster one; matched to the difficulty of the job, not applied globally.