Deep Research

---
name: deep-research
description: "Run a thorough multi-source research investigation end to end: plan, search broadly, verify, and synthesize a cited answer. Use when a question needs deep, autonomous research across many sources, not a quick lookup."
---

# Deep research

Deep research is the full loop from a hard question to a trustworthy,
cited answer: plan the investigation, gather from many sources, verify
what you find, and synthesize it into something more than the sum of the
pages. It is the method an agent (or a person) follows to research
autonomously without cutting corners.

## Method

1. **Plan before searching.** Sharpen the question, decompose it into
   answerable sub-questions, and decide what evidence would settle each and
   where it lives (see research-planning). A deep investigation without a
   plan becomes a pile of tangents.
2. **Search broadly, from multiple angles.** Cover each sub-question from
   several search angles and source types, so you are not captured by the
   first framing or the most popular result (see web-research). Cast wide
   first (find the landscape, the key sources, the disagreements), then go
   deep on what matters. Deliberately seek disconfirming evidence and the
   strongest opposing view, not just support.
3. **Go deep on primary sources.** Read the actual papers, docs, data, and
   original statements for the load-bearing claims, not summaries of them
   (see reading-papers, source-evaluation). The depth that distinguishes
   deep research is following claims to their origin and understanding
   them, not skimming aggregators.
4. **Verify as you go.** Evaluate every source and corroborate every
   important claim across independent origins before it enters your
   conclusion (see source-evaluation, fact-checking). Do not let an
   unverified claim propagate into the synthesis; mark what is confirmed,
   what is contested, and what is unknown.
5. **Synthesize, do not just collect.** The value is in connecting the
   findings: reconciling or surfacing the disagreements between sources,
   noting the consensus and the outliers, and answering the actual
   question rather than dumping what you found (see research-synthesis). A
   list of quotes is not research; the reasoned answer that accounts for
   them is.
6. **Deliver with citations and honest confidence.** Present the answer
   with its sources so it can be checked, distinguish well-supported
   conclusions from tentative ones, and state the open questions and what
   would resolve them (see writing-with-evidence). Know when to stop: when
   new sources stop changing the answer (saturation) or the decision is
   adequately supported.

## Boundaries

- Deep research is expensive; reserve it for questions whose stakes
  justify it. Most questions need a quick lookup and a source check, not
  the full loop (see research-planning's depth budget).
- Thoroughness is not the appearance of it: many low-quality sources are
  not depth, and a long report is not a good one. Depth is verified primary
  evidence and real synthesis, not volume (see the completeness-vs-padding
  tension in research-synthesis).
- For an autonomous agent, the loop must include verification and
  grounding at every step; an agent that searches and synthesizes without
  fact-checking confidently launders errors into a citation-shaped answer
  (see fact-checking, llm-guardrails).