Why spawn a subagent at all
Lesson 1 of 5 in Subagents and Context Isolation: The Delegation Move.
The multi-agent module gave you the honest reason to add a second agent: another context window. This module is that reason turned into a working skill — the delegation move. A parent agent spawns a subagent, hands it a brief, and gets back a compressed report. Done well, it is the single highest-leverage act of context engineering an agent can perform. Done carelessly, it is how good runs go quietly wrong.
Start with the accounting. The parent’s context is not just finite — it is load-bearing. It holds the user’s goal, the constraints discovered so far, the plan, the half-finished reasoning. Every token of bulk material you pour into that window sits in front of every decision the parent makes from now on. Context is the scarce resource; everything else in this module follows from that sentence.
Make it concrete. The task needs someone to read fifty source files to answer one question: which of these route handlers is missing an auth check? Read them inline and roughly 80k tokens of file contents move into the parent’s window — and never leave. Every later step now reasons past fifty files of noise, and the run degrades in the familiar ways: vaguer answers, re-fetching things it already saw, forgetting the user’s constraint from turn two.
A subagent can read all fifty files and return one paragraph. Its window fills with the same 80k tokens of bulk — and then the window is thrown away. The finding comes home; the mess does not. The subagent’s context is disposable in a way the parent’s never is: it gets garbage-collected the moment the report returns.
| Inline — the parent reads | Delegated — a subagent reads | |
|---|---|---|
What the parent retains | ~80k tokens of raw file contents, permanently occupying the window | One paragraph: the finding, plus file paths as evidence |
What later steps see | Every subsequent decision reasons past fifty files of noise | A clean window still dominated by the goal and the plan |
Total tokens spent | Lower — no second system prompt, no brief, no report | Higher — a fresh context pays its own setup, then does the same reading |
When it wins | The parent needs the raw contents verbatim — it is about to edit those files | The parent needs the answer, not the material |
Key terms: subagent, context window, context isolation, context engineering, token
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.