Member-only story
Chapter 1 covered the single agentic loop: message in, model decides, tool runs, result appended, repeat. That loop is the atom. This chapter covers the molecule — what happens when one agent isn’t enough, and you need a coordinator directing several of them.
This maps to Scenario 3 on the exam, the Multi-Agent Research System: a coordinator delegates to specialist subagents — one searches, one analyzes, one synthesizes — and you’re tested on orchestration, context passing, and handling partial failures gracefully. That’s the exact structure of this chapter.
Why split one agent into many #
A single agent with every tool bolted on sounds simpler, but it breaks down for three reasons that show up constantly on the exam:
Context pollution. A research task might involve dozens of search results, several rounds of analysis, and a long back-and-forth before a report gets written. If all of that lives in one conversation, the model is reasoning over an increasingly cluttered context window by the time it needs to write the final report — irrelevant search noise competes with the signal that actually matters.
Tool confusion. Chapter 1’s loop assumed the model always picks the right tool. That assumption gets shakier as the toolset grows. A search tool, a code execution tool, and a report-formatting tool sitting in the same call risk being called in the wrong order or for…