Sub-agent

A sub-agent is a second agent loop, started by the first, with an empty context window and a narrow brief. It runs its own turns, calls its own tools, and hands back a short answer. The parent never sees what it read to get there.

The reason is arithmetic. The window is finite and every token in it is re-read on every call. A search that opens thirty files to find one answer would spend the parent’s whole budget on material it will never need again. Run that search in a child and the parent pays for the conclusion instead of the transcript.

The cost: the parent cannot audit what it did not see, so a confidently wrong summary becomes the parent’s error with no way to notice. Brief a sub-agent the way you would brief a person, with what done looks like, what to return, and what to skip. A vague brief produces a vague summary, and the parent then reasons on top of it.