Deep Research Agent

A deep research agent is an AI agent that investigates a complex question by planning searches, gathering information from multiple sources, evaluating evidence, and synthesizing a cited report. It is designed for tasks that require more than retrieving one document or producing an answer from existing model knowledge.

A research run commonly includes:

• clarifying the question and decomposing it into subtopics;
• searching the web, private files, or databases;
• opening and extracting information from relevant sources;
• comparing claims and resolving contradictions;
• revising the search plan as new evidence appears; and
• generating a structured answer with citations and uncertainty.

Deep research is an agentic workflow, not simply a long prompt. The agent uses repeated tool calls and may consume substantial inference-time compute before producing its final output.

Source handling is the core quality problem. The agent must distinguish primary from secondary sources, account for publication and event dates, avoid circular citation, and ensure that each citation supports the specific claim attached to it. A polished report can still be wrong if retrieval or attribution is weak.

Production systems should preserve a trace of search queries, visited sources, extracted evidence, and discarded claims. This supports auditability and agent evaluation. Untrusted webpages also create indirect prompt injection risk.

Deep research agents are useful for literature reviews, market analysis, technical comparisons, due diligence, and policy research. They should not be treated as authoritative without checking high-impact conclusions against original sources.

The OpenAI deep research guide describes agentic search and synthesis with web search, file search, and remote MCP sources.