Exam Scenarios/Practice Scenarios
Advanced15 min

Scenario: Multi-Agent Research System

Design a coordinator-subagent research system that gathers, synthesizes, and validates information from multiple sources. Covers decomposition strategies, error propagation, and tool scoping.

Quick Reference

  • Coordinator decomposition quality directly determines report coverage — narrow decomposition = narrow reports
  • Structured error context (error type, scope, partial results) > raw exception propagation
  • Give subagents only the tools they need — a synthesis agent should not have web_search
  • Scoped tools (verify_fact) prevent subagent scope creep while enabling necessary validation
  • Inter-agent context should use structured formats (JSON, typed objects) not free-form prose

Scenario Description

You are designing a multi-agent research system that produces comprehensive, cited reports on complex topics. The system uses a coordinator agent that receives a research query, decomposes it into subtasks, and delegates those subtasks to specialized subagents. The subagents include: a web search agent (performs web searches and extracts key information), a document analysis agent (reads and summarizes long documents), a synthesis agent (combines findings from multiple sources into coherent narratives), and a report generation agent (formats the final output with citations and structure).

The system is built on the Claude Agent SDK using the subagent spawning pattern. Each subagent has its own system prompt, tool set, and context window. The coordinator communicates with subagents by spawning them with specific instructions and receiving structured results. Key challenges include: ensuring comprehensive topic coverage (the coordinator must decompose broadly enough), handling errors gracefully when a subagent fails (timeouts, API rate limits), and preventing subagents from exceeding their designated scope.