Cowork
Subagents
Delegation with its own context window. Used well, a subagent keeps expensive exploration out of your main thread. Used badly, it runs the same discovery three times in parallel.
This page is part of the Cowork hub, which covers the whole topic end to end.
A subagent is a delegated task that runs in its own context window and reports a result back to the main one. The parent gets the conclusion, not the hundred file reads it took to reach it — which is the entire point, and worth understanding before you use them, alongside the rest of the Cowork hub.
Why isolation saves tokens
Searching a codebase for where a behaviour lives might take forty file reads. Done in your main thread, all forty stay in context and are re-sent on every subsequent turn for the rest of the session. Done in a subagent, the forty reads are discarded and the parent receives three sentences.
| Approach | Context carried afterwards | Effect on the rest of the session |
|---|---|---|
| Search in the main thread | All 40 file reads | Every later turn pays for them |
| Search in a subagent | A short summary | Later turns stay cheap |
When subagents cost more
- Duplicate discovery. Three subagents each independently reading the same project structure triples that cost.
- Vague briefs. A subagent with an unclear goal explores until it runs out of room, then returns something the parent must redo.
- Trivial work. Spawning a subagent to rename a variable costs more in overhead than doing it inline.
- Verbose returns. A subagent that returns its full reasoning has defeated the isolation it existed to provide.
Scoping one properly
- Give it a single question with a checkable answer.
- Tell it where to look, so it does not spend calls discovering the layout.
- Specify the return format and keep it short — a list, a file path, a verdict.
- Do not chain more than you can reason about; nested delegation gets expensive fast.
The honest summary
Subagents are a context-management tool, not a speed tool. Reach for one when a task would otherwise pollute your main context with material you will never need again. If the work is small, or the answer is short anyway, do it inline.
The broader habits are in the optimization playbook, and the mechanics in tokens and context.