CCAR-F scenarios › Multi-Agent Research System · Question 25 of 60
Scenario 3 · Multi-Agent Research System
You're building a research system on the Claude Agent SDK in which a coordinator agent delegates work to specialized subagents: one searches the web, one analyzes documents, one synthesizes findings, and one generates reports. Given a topic, the system produces a thorough report with citations.
Tool Design & MCP Integration
To keep the configuration simple, every subagent's AgentDefinition grants the same 16 tools. The report generation subagent now sometimes calls web search and document-loading tools while formatting, which brings in unvetted material, and its formatting tool calls have become less reliable. What is the best fix?
Show answer and explanation
Answer: A — Restrict the report generation subagent's tools to the formatting and output tools its role needs, and give the other subagents tool sets scoped to their own roles in the same way.
Too many tools degrade selection reliability, and agents with tools outside their specialization tend to misuse them, so each subagent should get only the tools its role needs (task 2.3). A prompt-level prohibition (b) leaves the misuse possible and keeps the decision overhead of 16 tools. Merging agents (c) concentrates the same tool sprawl in fewer places. Discarding results after the fact (d) still wastes the calls and patches over a configuration problem it doesn't address.
More from this scenario
- The document analysis subagent has one tool, process_document(url, mode), described as…
- One of eight sources the web search subagent is retrieving returns an HTTP 503 once. The…
- Two credible sources give different figures for the same metric: an industry association…
- The coordinator concatenates outputs from six subagents, about 60 pages, and passes them…
- Final reports often contain correct claims with missing or wrong citations. Traces show…