Subagents: a fresh Claude with a job

The first way to add help is to stop doing a self-contained job yourself and hand it to another Claude.

A subagent is a second Claude you hand a job to. It goes off, does the work, and comes back with an answer — not the mess it made getting there.

Think of asking a colleague: “Go find out which files touch the login flow, and tell me.” They go read thirty files. They come back and say “these four.” You never see the twenty-six they ruled out — and you don’t want to.

Why “just the answer” is the whole point

Remember from Lesson 2.5 that a conversation is Claude’s working memory — everything it reads sits there in front of you.

So when you run a big search in your main conversation, all thirty files pile into that memory. Do that a few times and the conversation is bloated, slow, and full of detail you stopped caring about ten minutes ago.

A subagent does the reading in its own separate conversation. When it finishes, that conversation is thrown away and only the summary comes back to you.

That’s the part that trips people up: a subagent returns a summary, not its work.

The digging happened. You just didn’t have to watch it, and your conversation never had to hold it.

What it looks like

You say: “Search the codebase and tell me everywhere we read from the old config file.”

Claude spins up a subagent. Behind the scenes the subagent runs searches, opens files, follows references — dozens of small steps. You see a brief note that it’s working. Then one tidy answer:

“Three places: the startup script, the migration runner, and the health check.”

It felt like one step to you. It was dozens of steps for the subagent — just not in your context.

What’s next

Subagents are powerful, but they cut both ways: you can’t see their reasoning, and they can’t see your conversation. Next — which jobs to hand off, and which to keep for yourself.