For a broad topic, the coordinator delegates to four web search subagents covering different subtopics. Traces show it emits one Task call, waits for that result, then emits the next on a later turn. End-to-end latency is roughly four times the slowest single search. The searches are independent of each other. What change reduces latency most directly?

Answer: C — Have the coordinator emit all four Task tool calls in a single response, so the independent web search subagents run in parallel rather than one after another.

Parallel subagents are spawned by emitting several Task tool calls in one coordinator response rather than across separate turns (task 1.3). Collapsing the subtopics into one query (a) gives up the scope partitioning that ensures coverage, and the result would be shallower. A faster coordinator model (b) trims only the small gaps between calls while the subagent work stays serialized. Capping results (d) degrades research quality and still runs the searches one after another.