Summary: The trader

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

A decision is not yet an action: a separate agent takes the judge’s verdict and turns it into a concrete plan, building on the decision rather than reopening it, and it does so on the cheaper model. This lesson opens the code of the open-source TradingAgents framework (anchored to one frozen snapshot, marked 7e9e7b8) and reads the trader. This summary is the scan-in-five-minutes version of the full lesson.

Core ideas

• Deciding and doing are different jobs. Ask one agent to both decide and lay out the plan and it tends to re-argue while it should be executing. The system splits them: the judge decides, the trader operationalizes.
• The trader builds on the decision. It reads the judge’s plan straight from shared state and is instructed to anchor its reasoning on that plan and the analyst reports, not to reopen the debate. Its starting point is the finished judgment.
• Its output is a concrete proposal in shared state. The trader fills in a structured proposal (with a plain-text fallback) and writes it back as its own field, the handoff to the next stage. Same shape as the judge’s structured output, different job.
• The model tier follows difficulty, not importance. The trader runs on the quick (cheaper) model, while the judge before it ran on the deep one. The hard judgment already happened upstream, so the expensive model was spent there; turning a settled decision into a plan is real work but not the hard part.

What changes for you

When you chain AI steps, you keep hitting one fork: should a single agent both decide and produce the deliverable, or should those be two steps? Make them two. Let one step commit to the decision, then hand it to a second step whose only job is to turn that decision into the concrete artifact (the plan, the draft, the spec), feeding the decision in as settled so it builds on the call instead of relitigating it. And match the model to the difficulty of the thinking, not the importance of the output: spend your most capable model where the hard judgment happens, and let the operationalizing step run on something cheaper.