The multi-agent research system scenario puts you in charge of a coordinator agent that splits a research question across subagents, each exploring part of the problem, and then assembles a report. It draws mainly on agentic architecture and context management: how work is divided, how results flow back, and how the system stays accurate as information piles up.
The system you should picture
A user asks a broad question — the state of a market, a technical comparison, a literature review. The coordinator decomposes it into sub-questions, writes a brief for each subagent, and dispatches them. Each subagent searches and reads in its own context window, then returns findings to the coordinator, which reconciles them and writes the final answer with citations.
Decisions the exam probes
Is multi-agent worth it? Several agents cost more tokens and add coordination overhead. They earn that cost when the work is broad, parallelisable and too large for one context window. A short task with one tool does not need them.
The brief. Subagents start without the coordinator's context, so delegation has to say what success looks like: the objective, the scope boundaries, which sources to use, the format of the result and when to stop. Vague briefs produce vague findings.
What comes back. Subagents should return compact, structured findings rather than raw transcripts, so the coordinator's context does not overflow. Each claim travels with its source and the location of the supporting passage, so the final report can be verified.
Parallel or sequential. Independent sub-questions run in parallel to cut latency; steps that depend on earlier results run in order.
Conflicts and drift. When subagents disagree, both claims are kept with their sources and either checked further or reported as unresolved. Over long sessions, a compact record of established findings keeps early results from being contradicted later.
Common traps
- Returning full search transcripts to the coordinator.
- Dropping citations at the subagent boundary.
- Averaging two conflicting numbers into one that nobody reported.
- Running independent subagents one after another.
- Fixing vague subagent output with "be thorough" instead of a real brief.
How to practise
Write a coordinator that sends two subagents to research different halves of a question and requires each to return a list of claims with sources. Then make the two disagree on purpose and decide how your coordinator should handle it. Measure how much context the coordinator uses with raw transcripts versus structured findings — the difference is usually dramatic.
The six scenarios are listed in Anthropic's official exam guide (version 1.1), with four drawn per sitting. The emphasis described here is our own study guidance, not official exam content.