You are building an internal research assistant that answers complex analyst questions by coordinating multiple agents: one plans the task, one retrieves internal documents, one queries approved external sources, and one synthesizes a final answer. Users ask multi-step questions that often require comparing policies, summarizing recent changes, and citing evidence. The system is expected to support roughly 8,000 queries per day, with noticeable spikes during incident reviews and quarterly planning.
How would you design the agentic workflow and multi-agent orchestration for this system so it remains grounded, safe, and cost-effective under these constraints? Explain how you would decide when to use multiple agents versus a simpler flow, and how you would evaluate whether the orchestration is actually helping.