Five subagents in parallel: what you gain, what you pay
Each subagent starts from a fresh context: it rereads the files and returns only a summary. Time saved is capped at 1 / (1 − p), rereads are multiplied by N, and resending the context on every request (cache reads at 0.1 times input) shifts the balance. A calculator with editable assumptions compares one agent against N subagents.