Case Study: Design a Deep Research Agent

"Design a 'deep research' agent — something that takes a broad question and comes back in minutes with a structured, cited report, not a paragraph" is a different design problem from case-study-ask-the-web-agent's, even though both start with a question and end with citations. The web agent optimizes for a few seconds and one answer; a deep research agent is explicitly allowed 5–20 minutes, decomposes a question into several sub-questions, runs real multi-agent work on each, and has to survive a much longer failure surface: redundant findings across sub-agents, a synthesis step that has to reconcile disagreement rather than pick one source, and a citation contract that has to hold at the scale of a multi-page report instead of one paragraph. Treat it as three stages with genuinely different jobs — clarify and plan, execute in parallel, synthesize and cite — and the model-tier decision as a first-class design choice at every one of those stages, not an afterthought.

This subject is a model answer. It assumes you've read multi-agent-orchestration for the supervisor/worker mechanics this design applies rather than re-derives, reasoning-models-and-inference-time-scaling for what a reasoning model actually buys and when it's worth its cost, agent-architectures-and-the-agentic-loop for termination and budget controls, and evaluating-agents-and-multi-agent-systems for the trajectory/multi-agent evaluation machinery this subject's evaluation section applies to a research report specifically. Where a number appears, it is an assumption to state, not a fact to memorize.