---
id: con-deep-research
schema: concept/1
layer: garden
status: established
name: Deep research
aka: ["Research mode", "DeepSearch", "research agent", "Deep Research"]
summary: >
  An AI mode that spends several minutes searching many sources, reading and cross-checking them,
  and returns a longer cited report on a question.
first_seen: "2024-12"
relations:
  builds_on: [con-agents]
  evolved_from: []
  related: [con-verification, con-hallucination]
  parallels:
    - idea: "a literature review"
      field: "academia"
      note: "Gathering and summarising what many sources say on a question, with references a reader can follow."
    - idea: "an analyst's briefing memo"
      field: "consulting"
implementations:
  - { provider: anthropic, name: "Claude Research" }
  - { provider: openai, name: "ChatGPT deep research" }
  - { provider: google, name: "Gemini Deep Research" }
  - { provider: perplexity, name: "Perplexity Research" }
  - { provider: xai, name: "Grok DeepSearch" }
competencies: [cmp-frame-the-outcome, cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Every major provider ships one under a different name; a strong cross-provider tie with a clear verification lesson."
---

Deep research is a research agent. You ask a question, it may ask a few clarifying questions, and then it runs dozens of searches, reads pages and documents, and writes a report with citations. It takes minutes instead of seconds, and the result is much longer than a normal answer.

It is good for getting oriented fast: the range of options, the main arguments on a policy question, what competitors offer. It is weaker on very recent events, paywalled sources and niche facts, and it can cite a source that doesn't say what the report claims.

The report looks finished, which makes it easy to trust too much. Treat it as a well-organised first pass. Open the citations behind the claims you plan to rely on, and notice which important sources are missing.

## Example

A policy analyst asks for a summary of how five cities fund after-school programs. The report is useful, but two of the figures trace back to a 2019 page. She updates them from current budgets before sharing.

## How it connects

- A specialised **agent** focused on searching and reading.
- Many versions start with **clarifying questions** before the long job begins.
- Polished reports still need **verification**; citations reduce **hallucination** but don't remove it.

## Try it

Pick a question for deep research and write down three sources or facts a good answer should include. When the report arrives, check whether they're there, and open one citation to confirm it says what the report claims.