Design a "deep research" agent that produces a cited report on any topic.

Design details

1. Planning: break the question into 3–8 sub-questions; scale effort to the query (a simple fact needs 1 agent, a market analysis needs 5+). Without explicit effort scaling, agents over-research simple queries.
2. Sub-agent briefs: objective, scope boundaries, preferred source types, output format (claims + URLs + quotes), and a tool-call budget.
3. Search strategy: start broad, then narrow; prefer primary sources; record dates for freshness.
4. Findings store: keep findings outside the lead's context (files or a DB). The lead reads summaries, which keeps its context manageable over long runs.
5. Synthesis: resolve conflicting sources explicitly ("Source A says X; B says Y; B is more recent").
6. Citations: every claim links to a source. The verification pass checks that the source actually supports the claim, which catches hallucinated citations.
7. Durability: runs take minutes, so use checkpoints and resume. Return asynchronously with progress updates.

Costs and trade-offs: multi-agent research uses many more tokens than chat. It's worth it for high-value queries, not for "what's the capital of France". Add a router.

Evaluation: rubric-based judging (accuracy, completeness, source quality, citation correctness), checks against a set of known-answer questions, and human expert review for a sample.

Related

• Previous: Q37. Design a coding agent like Claude Code or Cursor's agent mode.
• Next: Q39. How do you design authorization for an agent that acts on behalf of users across systems (email, CRM, drive)?
• Deep Agents
• LangGraph