Core concept: In hub-and-spoke, the coordinator spawns workers with the subagent tool. That tool must be allowed. Each subagent starts on a clean context. Parallelism comes from one response carrying several of those calls.

Why A and D are correct:

A: the SDK docs say "Claude invokes subagents through the Agent tool, so include Agent in allowedTools". Without it, "Agent invocations fall through to your canUseTool callback or, in dontAsk mode, are denied". The tool was "renamed from Task to Agent" in later versions, so the rule is the same under either name.
D: "multiple subagents can run concurrently, so independent subtasks finish in the time of the slowest one rather than the sum of all of them." The lead agent "spawns subagents to explore different aspects simultaneously".

Why the others are wrong:

B: the docs say the opposite. A subagent's "context window starts fresh, with no parent conversation", and it does not receive "the parent's conversation history or tool results". You pass what it needs in the tool's prompt string.
C: one assistant response can carry several tool_use blocks. That is exactly how the fan-out works, so C also contradicts D.

Remember: Allow the subagent tool, hand context over in the prompt, and fan out in a single response.

Question 164