Lesson 24 — Multi-agent: planner and worker

Two agents, two jobs

A multi-agent system splits work between several agents, each with its own system message and its own short message list:

• the planner breaks the goal into small tasks, and later combines the results;
• the worker does exactly one task at a time.

Your code is the orchestrator: it passes the planner's tasks to the worker, collects the answers and sends them back to the planner.

goal -> planner -> [task, task, task] -> worker x3 -> results -> planner -> final

Here both "agents" are fake functions; with a real API they would often be the same model called with different system messages.

Why it matters: small focused prompts are easier to get right than one giant prompt. The cost: more calls, and more places for mistakes — so only split when one agent struggles.