Research Agent
A research agent is an AI system that answers an open-ended question by running its own investigation: it plans what to look up, issues searches, opens and reads sources, notes what is still missing, searches again and finally writes a structured report with citations. The output is closer to a briefing memo than a chat reply, and the process may take many minutes and consume dozens or hundreds of page reads. The difference from ordinary retrieval-augmented generation is control flow: classic RAG retrieves once and answers, whereas a research agent decides in a loop what to retrieve next based on what it has learned so far.
The core pattern is an iterative search-read-synthesize loop. The agent first decomposes the question into sub-questions and a rough outline; for each it formulates queries, fetches results, reads the pages or documents (often with a separate model call to extract the relevant passages), and records findings with their source. It then reflects: are there contradictions, gaps or weak sources? If so it issues refined queries, follows citations from one document to another, or switches strategy, and only when its stopping rule fires does it assemble the final write-up. Public research prototypes such as STORM from Stanford (Shao and colleagues, 2024), which generates Wikipedia-style articles through simulated multi-perspective question asking, and many open-source agentic search projects follow variants of this recipe.
Deep research became a mainstream product category in the winter of 2024 to 2025. Google introduced Deep Research in Gemini in December 2024, OpenAI launched a Deep Research feature in ChatGPT in February 2025, and Perplexity and others released comparable modes within weeks. Each combines a reasoning model, web browsing and long-form report generation, and vendors describe training or tuning the model specifically for browsing and analysis tasks; the details of the exact training and the reported benchmark results differ by vendor and are best taken from their own documentation. Browsing-focused benchmarks such as BrowseComp (introduced by OpenAI in 2025) were built to measure this kind of persistent, multi-step information seeking.
Citation checking is the part that separates a trustworthy research agent from an eloquent one. A report is only as good as the link between each claim and the passage that supports it, and models can produce a citation that looks right but points to a page that does not say what is claimed, or invent a source entirely. Good systems attach a claim to a specific retrieved span, verify afterward that the span entails the statement, prefer primary sources over aggregators and keep the raw links so a person can click through. Even so, a real URL is not proof of a supported claim, and a sentence summarizing several sources can blend them in ways none of them stated.
The pipeline has characteristic failure modes beyond hallucination. Search engines and the open web are biased toward what is popular, recent and search-optimized, so an agent can collect many pages that merely repeat one weak claim and treat the repetition as corroboration. Paywalled, scanned or non-English sources are often invisible, and content generated to rank well may itself be machine-written. Agents may stop too early, stay confident while evidence conflicts, or follow an irrelevant branch for many steps; hidden instructions in a fetched page can attempt prompt injection against the agent. A careful design therefore logs the queries and sources, limits the sites it will trust and acts with read-only permissions, since it needs to read widely but has little reason to write anywhere.
A common misconception is that more searching automatically means more accuracy, and that a long cited report is a verified one. Length and formatting are persuasive cues, yet they say nothing about whether sources were read correctly or whether the question was framed well. Research agents work best as fast first-pass assistants for literature scans, market overviews, background reading and comparison tables, where an expert then audits the key claims against the sources. For decisions with legal, medical or financial stakes, the report is a lead list for human verification rather than an authority, and the questions it handles worst are the ones where the true answer is that experts disagree or the evidence is thin.
Read research agent as retrieval turned into a loop: a model that plans, searches, reads and rewrites until its own gap check is satisfied, whose usefulness rests on how well each sentence is tied to a source that really supports it, and whose polished report is a well-organized starting point for a human to verify rather than a finished verdict.