Subagents

Delegation here solves a space problem, not a capability one. Your main Claude could read those forty files itself — but doing so would fill the conversation with forty files’ worth of material, and everything that follows would be reasoned about through that clutter. Long threads get worse, not better, and the usual culprit is the messy middle of some side-quest nobody needed to see.

A subagent absorbs that middle somewhere else. It gets its own fresh context window, does the digging, and hands back the paragraph you actually wanted. The forty files were read; they just never entered your thread. That’s why the classic uses are all shaped like research: find every place we do X, read this whole folder and tell me the five things that matter, check whether this pattern appears anywhere else.

The trade-off is the one rule people trip on. A subagent cannot see your conversation. It starts blank, so every constraint, preference and piece of background that made the request make sense has to be stated inside the request itself. A vague delegation comes back with a confident, well-written answer to a question you didn’t ask.

You mostly don’t configure any of this. In Claude Desktop you ask in plain language — “use a subagent to…” — and Claude spins one up when it helps. The /agents command below is for the further step of defining named specialists you reuse, which is a team-setup move rather than an everyday one.

Where this goes next is agent teams — several subagents working at once rather than one helper at a time. The comparison below is worth reading first, because “agent” and “subagent” get used interchangeably in conversation and they are not the same thing.