Summary: Why split one AI into many

For education only. This teaches multi-agent architecture using stock analysis as the example; it is not investment, financial, or trading advice.

A real agent team is one workflow split into specialist roles by function, with your strongest model spent only where one agent decides for the rest. This lesson uses the open-source TradingAgents framework (anchored to one frozen snapshot, marked 7e9e7b8) to make that concrete. This summary is the scan-in-five-minutes version of the full lesson.

Core ideas

• The team is a decomposition, not a headcount. The framework groups its agents by function: four analysts who gather information, two researchers who argue opposite sides (a bull and a bear), one trader who turns the argument into a plan, three risk reviewers who stress-test it, and two managers who judge. The groups follow the natural steps of one workflow. With all four analysts on, that is twelve agents, but the number is not the point.
• Each role is a small, focused job. Splitting buys focus (a narrow agent with a narrow job is more reliable than a generalist), deliberate opposition (separate bull and bear agents argue at full strength instead of one model talking itself into a tidy answer), and replaceability (each role is built separately, so you can swap one without touching the others).
• The work flows in a clear order. Analysts gather, then the bull and bear debate, then the research manager rules, then the trader plans, then the three risk reviewers stress-test, then the portfolio manager makes the final call. The two judges sit at two different points in the flow.
• The decision most people miss: spend capability at the judgment points. The system runs on two model tiers, a more capable, costlier model and a cheaper, faster one, and it gives the capable model to exactly the two judges (the research manager and the portfolio manager). Everything else runs on the faster model. Change how many analysts you run and the headcount shifts, but the capable model still goes to exactly those two judges. That steady rule, not the number twelve, is the design.

What changes for you

Before this lesson, “split one AI into many” was a slogan. Now you can look at a real agent team and see the two decisions underneath it: break the workflow into roles by function, then find the small number of points where one agent commits on behalf of the rest and put your best model only there. That is a pattern you can carry into any agent system you design, in any domain. It is also how you judge what an AI hands you: good structure in, good results out.