Agentic research design lessons
Lessons distilled from production deep-research agents — GPT-Researcher, Stanford STORM, LangChain and HuggingFace Open Deep Research, OpenAI and Gemini Deep Research, and Perplexity.
Orchestration
- The orchestrator plans and synthesizes; it never does ground research itself.
- Budget per-subtopic specialists with isolated context.
- Outline before searching; reflect at both supervisor and worker level.
- Stop on coverage (two independent sources, or one authoritative one), not on a loop count.
Sourcing and retrieval
- Multi-engine, domain-routed sources; rerank and quality-gate before the LLM sees them.
- Filter by freshness and authority / venue reputation.
- Read full pages, not snippets.
Citation and provenance
- Attach citation metadata at retrieval time, not after generation — pre-embedding citations gave the best observed accuracy (~92% in Perplexity).
- Cite at the claim / sentence level with the supporting quote.
- Keep stable, numbered citations through summarization.
Verification
- A dedicated reviewer with explicit acceptance criteria.
- Cross-source corroboration is the primary anti-hallucination defense.
- Check URL health / citation existence — even top tools hallucinate 3-13% of URLs.
- Perspective diversity is structural anti-hallucination.
Reporting
- Outline-driven report with a confidence / coverage indicator per section.
- Keep intermediate artifacts (outline, mind-map) for human review.