Multi-Agent Research System

Why it matters

A lead-and-subagents design is easy to describe and hard to observe. This demo shows the contracts, the tool boundaries and the tracing that make one inspectable.

What to practise

Run the lead agent, which splits a request into two to four subtopics and delegates through the subagent tool. In the code, each AgentDefinition gives a role its own tools: researchers get WebSearch and Write, the data analyst gets Glob, Read, Bash and Write, and the report writer adds Skill. Watch the PreToolUse and PostToolUse hooks attribute every tool call to a subagent through parent_tool_use_id, and read the transcript and tool_calls.jsonl they write. The demo’s lead allows the tool under its earlier name, Task; the current Claude Agent SDK calls it Agent.

How it fits

The parallel researchers are the case 1.4.1 asks you to justify on parallelism and context isolation. The lead-controlled graph and its tracing are one topology from 1.4.2. The per-role prompts and tool lists are 1.4.3’s delegation contracts, and the notes the researchers save under files/research_notes/ are the evidence the analyst and the report writer accept.