Subagents

A subagent is a separate Claude you delegate a job to — "go read these forty files and tell me where logins are handled." It works in its own context window and reports back just the answer, so your main chat doesn't fill up with the messy middle. Great for research, search, and any side-quest.

The reason a subagent helps isn’t just the extra pair of hands — it’s the fresh context. A big research dig (“read these forty files and tell me where logins are handled”) generates a mountain of intermediate reading. Done in your main chat, all of that piles up and crowds out the thread you actually care about. Hand it to a subagent and that mess stays in its window; you get back only the answer.

The trade-off is that a subagent can’t see your conversation. It starts clean, so whatever it needs to know has to be in the request you give it. That’s a feature as often as a limitation — it means the helper isn’t distracted by everything else you’ve been discussing — but it does mean a vague hand-off gets a vague result. Spell out the goal and the context up front.

Reach for one when a task is a self-contained side-quest: searching, reading widely, checking one specific thing. When the work is wide — many independent pieces at once — that’s when subagents scale up into agent teams, several running in parallel and reporting back together.

Use a subagent to find every place we send email, and summarize how each one works.

Spin up a subagent to read the whole /docs folder and come back with the 5 things I most need to know.