Core concept: Multi-agent systems use a hub-and-spoke shape. One coordinator delegates the work, and every result travels back through it. Subagents do not talk to each other and do not share memory.

Why A and C are correct:

Anthropic's research system uses "a lead agent" that "coordinates the process while delegating to specialized subagents that operate in parallel".
Parallelism comes from one coordinator response carrying several Task (now Agent) tool calls. Split them across turns and they run one after the other.
The Agent SDK lists "Parallelization: multiple subagents can run concurrently, so independent subtasks finish in the time of the slowest one".
Routing every hand-off through the coordinator leaves one place to trace the run and to handle failures the same way each time. Anthropic added "full production tracing" for exactly that reason.

Why the others are wrong:

B: A flat mesh with shared global state removes that single point of control. The docs describe the lead agent as the hub, not a peer in a shared pool.
D: Subagents inherit nothing. "Each subagent starts with a fresh, isolated context window", and the parent's conversation history is explicitly listed as something it does not receive. The coordinator must put the needed facts in the subagent's prompt.

Remember: Coordinator in the middle, several Task calls in one response, fresh context for every subagent.

Question 142