WTF are subagents (or agents) in Claude Code?

What is a subagent?

You can think of Claude Code as having one main conversation thread — and this is what consumes tokens up to its context limit. But subagents (or "agents"; both terms are used synonymously in Claude Code) allow that main thread to spawn off additional threads, each with their own separate context windows.

This allows you to pass specific jobs off to subagents and have them run in their own processes. Once a subagent spins up, it runs autonomously — you can't interact with it mid-task.

Subagents are defined in markdown files with corresp...