Deep Research
An AI agent capability where the model autonomously conducts multi-step research on a topic—planning search queries, reading multiple sources, synthesizing findings, and producing a comprehensive report. Unlike single-query search, deep research agents iterate: they search, read results, identify gaps, refine their queries, and repeat until they've assembled a thorough analysis. The pattern emerged in 2025 with products like Gemini Deep Research, ChatGPT's research mode, and Perplexity's Pro Search, and is now a core capability expected of knowledge-work AI agents.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A venture capital analyst asks an AI agent to research a startup's competitive landscape. The agent plans 8 search queries, reads 25 sources, identifies 12 direct competitors, maps their pricing and positioning, finds 3 recent funding announcements, and produces a 2,000-word competitive brief with citations—work that would take a human analyst 4-6 hours completed in 15 minutes.
Frequently asked questions
- How is deep research different from regular AI search?
- Regular AI search answers one question with one or a few sources. Deep research is an agentic workflow: the model plans multiple search queries, reads and evaluates results, identifies information gaps, searches again to fill those gaps, cross-references across sources, and synthesizes everything into a structured report. The key difference is iteration—deep research agents make 5-20 search queries per task, not just one.
- What are the limitations of deep research agents?
- They can only research publicly available information (or information in connected knowledge bases). They may miss very recent events if search indexes haven't caught up. They can occasionally synthesize conflicting sources incorrectly. And they consume significantly more tokens than single-query interactions—a deep research task might use 50,000-200,000 tokens, costing $0.50-$5 per report depending on the model.