What is agentic research?
Agentic research is research where an AI agent chooses each next step. You describe what you need to know; the agent breaks it into sub-questions, searches, reads, notices gaps, searches again and writes up what it found with citations. A chatbot with web search retrieves once and answers. A research agent keeps going until its plan or its budget is exhausted.
The category went mainstream in early 2025. Google shipped Deep Research in the Gemini app in December 2024, OpenAI launched deep research in ChatGPT in February 2025, Perplexity followed within two weeks, and Anthropic added Research to Claude in April 2025. In 2026 the products moved toward private data. OpenAI's February update put deep research on GPT-5.2 and let users restrict it to specific sites and connect apps; Google's April 2026 Deep Research and Deep Research Max agents accept MCP servers so they can query internal databases next to the web.
How does a research agent work?
Every product implements some version of the same loop. Anthropic described its own in detail in June 2025: a lead agent plans the research and starts subagents that search in parallel, and a separate citation agent matches claims to sources at the end (Anthropic Engineering). OpenAI says a deep research task takes 5 to 30 minutes.
| Stage | What the agent does | Typical failure | Control |
|---|---|---|---|
| Clarify and plan | Asks scoping questions, splits the question into sub-questions, proposes a plan | Plans around what is easy to search, not what the decision needs | Answer the scoping questions; review the plan where the product shows one |
| Search | Runs many queries, often through parallel subagents | Picks SEO-optimized pages over primary sources | Restrict to named sites or databases; prefer scholarly or internal corpora when they exist |
| Read | Opens pages, PDFs, files and connected apps; extracts claims | Paywalled or JavaScript-heavy sources are skipped silently | Upload the documents that matter instead of hoping it finds them |
| Iterate | Decides what is missing and searches again | Stops when the budget runs out, not when the question is answered | Ask for an explicit list of open questions and gaps |
| Write and cite | Synthesizes a report and attaches citations | Citation points to a page that does not support the sentence | Sample-check citations; trace every number and quote |
Two numbers from the same Anthropic write-up explain the economics. Its multi-agent setup beat a single agent by 90.2 percent on an internal research evaluation, and multi-agent systems used about 15 times the tokens of a chat. On the BrowseComp benchmark, token usage alone explained 80 percent of the variance in performance. Research quality is mostly bought with compute, which is why every vendor rations it by plan.
The same team reported that early versions preferred SEO-optimized content farms over academic PDFs and expert blogs until they added source-quality rules to the prompts. Source selection is a design choice, and you rarely get to see how a vendor made it.
How reliable are research agent citations?
Less reliable than the formatting suggests. The DeepTRACE audit (Venkit, Laban and colleagues, September 2025) tested deep research configurations of several public systems and found citation accuracy between 40 and 80 percent, large fractions of unsupported statements, and answers that were highly one-sided on debate questions. A March 2025 Tow Center study of eight AI search tools found wrong citations in more than 60 percent of its tests, and the paid tiers were not consistently better than the free ones. OpenAI's own launch post says deep research can hallucinate facts, make incorrect inferences and struggle to tell authoritative information from rumor.
The mechanism matters more than the percentages, which change with every model release. A research agent writes a sentence from what it has read, then attaches a source. The source can be on topic without supporting the specific claim: a different year, a different figure, a vendor's own marketing. Long reports compound the problem, because each section inherits the errors of the searches beneath it. We cover the general pattern in hallucinations at system scale.
The practical rule: a citation tells you where to look, not that the claim is true. Numbers, quotes, dates and anything that would embarrass you if wrong need a person to open the source.
Which kind of research agent fits which job?
The products split by where they search, not by how clever the model is. Pick the corpus first.
| Type | Products | Best for | Watch |
|---|---|---|---|
| General web research | ChatGPT deep research, Gemini Deep Research, Perplexity Deep Research | Market scans, vendor landscapes, policy and technology overviews | Web source quality; one-sided answers on contested topics |
| Work-data research | Claude Research (Google Workspace, connectors), Microsoft 365 Copilot Researcher | Questions whose answer is in email, docs, meetings and chats as well as on the web | Connector permissions; what the agent can read that the requester should not |
| Scholarly research | Elicit, Consensus, Undermind | Literature reviews, evidence checks, R&D and medical questions | Corpus coverage; preprints mixed with peer-reviewed work |
| Research APIs | Gemini Deep Research agent, OpenAI o3-deep-research, Perplexity Sonar Deep Research | Research embedded in your own product or workflow | Cost per task; you own evaluation and logging |
Scholarly tools deserve more attention than they get in companies. Elicit, Consensus and Undermind search paper databases rather than the open web, so a citation points to a study rather than a blog summarizing it. For R&D, medical, policy and technical due-diligence questions, that is a better starting point than a general agent.
Where do research agents fit in a company?
- First drafts of desk research. Market maps, competitor overviews, regulatory summaries, technology option papers. The agent produces the draft and the reading list; an analyst checks and cuts.
- Vendor and partner screening. A first pass on a shortlist: what a vendor sells, recent funding, public incidents. Useful for deciding where to spend diligence time, not as the diligence.
- Literature reviews. Scholarly agents find and summarize papers faster than keyword search. Elicit now supports systematic reviews built for PRISMA 2020, which shows where the tooling is heading.
- Internal synthesis. With connectors, an agent can pull together what the company already knows from email, docs and meeting notes before anyone searches outside.
Where they do not fit: final legal or medical advice, numbers going into a board deck or filing without a trace to source, and contested questions where a one-sided summary would mislead. Our general rule for when not to use AI agents applies: if a wrong answer is expensive and hard to detect, keep a person in the loop.
How should you govern research agents?
Govern by where the output goes, not by which tool produced it. The same report is harmless as personal reading and risky as a client deliverable.
| Tier | Use | Controls | Review |
|---|---|---|---|
| 1. Explore | Personal learning, first look at a topic | Approved tools only; no confidential data in consumer accounts | None beyond the user |
| 2. Inform | Input to an internal decision: vendor shortlist, market sizing, technical options | Business or enterprise plan; connectors scoped by role; report saved with its sources | Requester checks every number and quote that the decision depends on |
| 3. Publish | Anything external, regulated or contractual: client deliverables, filings, published content, clinical or legal claims | Restricted source lists or scholarly corpora; audit log of prompts and sources | Named reviewer traces each claim to a primary source and signs off |
Connectors are the control most teams skip. Anthropic's documentation says Claude Research uses any connectors enabled in the chat automatically, without asking for approval each time (Claude Help Center). Microsoft requires an admin to enable Researcher and separately approve Claude models (Microsoft Support). Decide which data sources each role may connect before rollout, and write it into your AI usage policy.
Should you build on a research API?
If research is part of your product or a repeated workflow, the APIs are mature enough to use. Google's Deep Research agent runs through the Interactions API in preview, with collaborative planning, MCP servers, File Search over your own documents, and background execution that typically takes several minutes. OpenAI added o3-deep-research and o4-mini-deep-research to its API in June 2025 with MCP support and webhooks for long-running tasks. Perplexity sells Sonar Deep Research by tokens plus search fees.
Buy the agent, own the evaluation. Log every source the agent read, score a sample of reports for citation accuracy against your own gold set, and set a cost ceiling per task. Our AI agent evaluation guide covers the method.
What to ask a vendor
- Can we restrict the agent to named sites, databases or internal sources for a given task?
- Which connectors can it read, and can admins scope them by role?
- Do you publish citation accuracy on a benchmark we can rerun, or only answer quality?
- Are prompts, uploaded files and connected data used for training, and how long are reports retained?
- How is usage metered: tasks per month, a shared compute pool, credits or tokens?
For scored vendor comparisons, see our partner site's Best AI Research Agents 2026. For the broader picture, see the agentic AI pillar.
Related questions
Frequently asked questions
What is agentic research?
Agentic research is research carried out by an AI agent that decides its own next step. Given a question, the agent plans a set of searches, runs them, reads the pages or papers it finds, follows up on gaps and writes a report with citations. Products include ChatGPT deep research, Gemini Deep Research, Claude Research, Perplexity Deep Research, Microsoft 365 Copilot Researcher, and scholarly tools such as Elicit, Consensus and Undermind.
What is the difference between AI search and a research agent?
AI search answers one query from one round of retrieval in seconds. A research agent runs many rounds: it breaks the question into parts, searches each, reads full sources, and changes its plan based on what it finds. That takes minutes, uses many times the compute, and produces a long report instead of a paragraph.
Are deep research citations reliable?
Not reliable enough to skip checking. The DeepTRACE audit (arXiv, September 2025) measured citation accuracy between 40 and 80 percent across deep research configurations and found large shares of unsupported statements. Treat each citation as a lead to verify, especially for numbers, quotes and claims that a decision depends on.
Which AI research agent is best for a company?
It depends on the data. For general web research, ChatGPT deep research, Gemini Deep Research and Perplexity are the main options. For research that has to include email, documents and meetings, Claude Research with Google Workspace or Microsoft 365 Copilot Researcher fit better. For scientific literature, Elicit, Consensus and Undermind search paper databases instead of the open web. Our partner site We The Flywheel scores eight of them.
Can you build your own research agent?
Yes. Google offers Deep Research and Deep Research Max agents through the Gemini Interactions API, with MCP and File Search for private data. OpenAI exposes o3-deep-research and o4-mini-deep-research in its API with MCP support and webhooks, and Perplexity sells Sonar Deep Research. Building on these is faster than writing your own orchestration, but you take on cost control and evaluation.
Sources
- Anthropic, How we built our multi-agent research system (13 June 2025); Anthropic, Claude takes research to new places (15 April 2025); Claude Help Center, Use research on Claude.
- Venkit et al., DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence, arXiv 2509.04499 (2 September 2025).
- Nieman Lab, AI search engines fail to produce accurate citations in over 60% of tests (March 2025).
- OpenAI, Introducing deep research (2 February 2025, updated 10 February 2026); Wikipedia, ChatGPT Deep Research; The Decoder, OpenAI expands API access to deep research models (26 June 2025).
- VentureBeat, Google's new Deep Research and Deep Research Max agents (21 April 2026); Google AI for Developers, Gemini Deep Research agent.
- Microsoft Support, Get started with Researcher in Microsoft 365 Copilot; Engadget, Perplexity has its own Deep Research tool now too (15 February 2025); Elicit, Blog (PRISMA 2020 systematic review, 6 May 2026).