Multi-Agent Research System

Why it matters

This demo is a working coordinator–subagent system on the Claude Agent SDK: 1.2’s decisions in running code.

What to practise

Run the lead agent, whose only tool is Task, an earlier name of the Agent SDK’s Agent tool. It splits a topic into two to four subtopics, runs researcher subagents in parallel, then a data analyst and a report writer. They hand off through files: researchers write to files/research_notes/, which the others read. Each subagent is an AgentDefinition with its own description, prompt and tools. Read the PreToolUse and PostToolUse hooks, which log every tool call and attribute it to a subagent through parent_tool_use_id.

How it fits

The lead agent is the coordinator of 1.2.1, and its subtopics are the partitioning of 1.2.2. Ask whether every request needs this fixed pipeline; 1.2.3 is the decision to compare against. The definitions are 1.3.1’s contracts, and the file handoff replaces the prompt-passed findings of 1.3.3.