When decomposition actually pays off
A subagent is a separate agent instance, with its own context window, dispatched to handle a piece of a larger task and report back a result rather than a full transcript. The appeal is straightforward: instead of one context accumulating everything, from raw file contents to noisy tool output to intermediate dead ends, that mess happens inside the subagent's own context, and only the distilled outcome crosses back into the main conversation. The orchestrating agent stays lean, and the total amount of work that could plausibly get done goes up, because independent pieces can proceed without waiting on each other.
That benefit is real but conditional: it shows up specifically when the pieces of work are genuinely independent, meaning one piece doesn't need to know what another piece is doing while it works, and when the intermediate output of a piece is large or noisy relative to the useful result that comes out of it. Searching five unrelated parts of a codebase for a pattern, or gathering background on several unrelated topics, are good candidates: each search can run without the others, and each produces a lot of scanning and false leads that shouldn't need to live in the main conversation, just a clean answer at the end.