How do you design a deep research agent?

Every deep research agent solves four problems. Every design choice maps back to one. Scoping. Vague question to a bounded investigation. Stopping. Knowing when the answer is good enough. Grounding. Every claim is tied to real evidence. Memory. Handling more content than one context window holds. The loop: Intake → Plan → Orchestrator (pick sub-question, call tool, note result, decide continue) → Compose → Verify → Stop. Pick a side on the fork. Single agent is cheaper and more coherent, weaker on breadth. Orchestrator with subagents is broader at roughly 15x the tokens. The consumer is single. Enterprise research earns the orchestrator. North Star: Weekly Trusted Research Tasks. No reports generated. That rewards confident wrong outputs.