Subagents are isolated Claude Code instances the main session can spawn for focused work. They have their own context, their own tools, and return a clean result when done.

What it does

Claude delegates a task to a subagent with a specific prompt and toolset. The subagent runs in isolation - its reads, edits, and reasoning don't land in the main context. When it finishes, it returns a summary. This is the building block for parallel work, research fan-outs, and context hygiene in long sessions.

When to use it

Research tasks that would otherwise swell the main context.
Parallel exploration across several files or topics.
Delegating a concrete subtask while you keep the bigger picture.
Running specialized roles (reviewer, auditor, researcher) with different system prompts.

Gotchas

Subagent output is what the main session sees - make sure the return is informative.
Each subagent is a separate model call. Spawning too many inflates cost and time.
Subagents don't share memory with the parent unless you configure persistent memory.

Official docs: https://code.claude.com/docs/en/sub-agents.md