Subagents Investigation — Claude Code Guide

A subagent compresses research

When you need to review many modules, tests, configurations, or logs, a subagent can move the search into a separate context and return only a concise, verifiable result to the main session.

A subagent is a research tool, not a replacement for the final decision. The main session is responsible for checking the conclusions and making changes in the repository.

A mini-task spec for a subagent

Investigate the refund operation path.
Check controllers, services, payment client, config, and tests.
Return a concise conclusion, file:line evidence,
assumptions, open questions, and next steps.
Do not change files; do not copy large code fragments.

Expected result

• **Conclusion** — a few sentences about the flow you found;
• **Evidence** — specific references to files and lines;
• **Assumptions** — what follows from indirect signals;
• **Open questions** — what remains unverified;
• **Next actions** — what the main session should check.

After the conclusion is returned

1. Manually check two or three key locations.
2. Compare the conclusion with a test or command.
3. Separate confirmed information from assumptions.
4. Update the inventory, API map, or investigation note through a normal diff.
5. Do not transfer an unverified summary into documentation as fact.

Delegation boundaries

Do not use a subagent for every small task or build a complex multi-agent system during a simple investigation. Delegation should reduce noise while preserving human control over the evidence and the final result.