A subagent is running concurrently while you keep working, and its answer is the thing your next instruction depends on. At what point can you act on what it found?

The principle — a concurrent worker's result has a delivery moment of its own, and it is later than the moment the work is done.

Why the key is correct — the two sentences fit together exactly. Background subagents run concurrently while you continue working, and a background subagent's results reach Claude as a completion notification in a later turn. Concurrency is what lets you keep going; the later turn is what you pay for it. So the earliest point you can act on the finding is a turn after the one you were in when it landed.

Why the others are wrong — finishing is not delivery: the worker can be done while the conversation has not yet been told. Nor does the notice cut into the turn in progress, which would defeat the whole reason for dispatching it that way. And a notification arrives rather than waiting to be collected, so there is nothing to poll for.

Remember this — plan the next step for the turn after, not the same one. If the answer is needed inside this turn, this is the wrong mode to have chosen.

Sources — Anthropic's Claude Code subagents documentation.