Question 14
Domain 2A coordinator agent needs to gather data from three independent sources (web, database, document store) before synthesizing a final answer. Currently it spawns each subagent in a separate turn — web search, then wait, then database, then wait, then document store, then wait — taking ~90 seconds total. How should you modify the coordinator to reduce latency?
Correct answer: B
Explanation
The coordinator should issue all three Task tool calls in one response because independent work can run concurrently instead of serially. Parallel execution removes the repeated wait time between turns, so the web, database, and document store lookups can finish at the same time before synthesis.
Why each option is right or wrong
A. Use fork_session so each subagent runs in its own isolated session
Session isolation separates context, but does not by itself make sequential turns run concurrently.
B. Emit all three Task tool calls in a single coordinator response; they will execute in parallel
Under the agent orchestration pattern, independent subtasks should be dispatched in the same assistant turn so the runtime can schedule them concurrently rather than forcing a turn-by-turn dependency chain. Here, the web, database, and document-store lookups have no interdependence, so emitting three Task tool calls together collapses the three separate wait cycles and reduces total latency from roughly 90 seconds to approximately the duration of the slowest single lookup.
C. Configure the coordinator's max_tokens higher so it can process all three in one reasoning pass
Higher token limits affect response length or reasoning budget, not external tool-call scheduling latency.
D. Chain the subagents so each passes its results directly to the next without returning to the coordinator
Chaining makes tasks sequential and only helps when later steps depend on earlier outputs.