How Deep Research Agents Work

Deep research products became widely available in late 2024 and 2025. Google introduced Gemini Deep Research in December 2024, OpenAI launched Deep Research in ChatGPT in February 2025, and Perplexity, Anthropic and others released similar research modes. They differ in details, but most follow the same loop. First, the agent interprets the request and often asks clarifying questions about scope, audience or time frame. It then drafts a research plan: sub-questions to answer and searches to run. Next comes the iterative phase. The agent issues searches, opens promising pages or PDFs, extracts relevant passages, and decides what to do next based on what it found. If a source mentions a new term or a conflicting figure, the agent can run follow-up searches. This loop can run for many minutes and touch dozens or hundreds of pages, the main difference from a normal chatbot answer that uses one quick search. Finally, the agent synthesizes. It organizes notes into sections, writes a report, and attaches citations to specific claims. Some systems use several sub-agents working on different sub-questions in parallel, with a lead agent combining results. The weaknesses are predictable. Agents can only read what they can reach, so paywalled journals, login-gated data and sites that block bots are often missing. They can over-trust search rankings, repeat SEO-driven content, or mix up dates and figures across sources. A citation next to a sentence does not guarantee the source actually says that, so spot-checking is necessary. Reports can also sound more certain than the evidence justifies. A useful mental model: a deep research agent is a fast, tireless research assistant that produces a solid first draft and a list of leads, not a final authority.