Many real-world questions cannot be answered responsibly from a single model call. They require searching across sources, reading the original material, separating evidence from interpretation, comparing conflicting claims, and producing an answer that a reader can verify.

Deep research agents are built for that kind of evidence-gathering work.

A deep research agent goes beyond retrieval. It plans sub-questions, searches the web or private corpora, reads sources, extracts claims, tracks provenance, identifies gaps, searches again, and synthesizes a report with citations. The agent's job is not to sound confident. Its job is to make the path from question to evidence auditable.

In this chapter, we will examine the search-read-synthesize loop, source reliability, contradiction handling, confidence, and the architecture behind research agents that can produce useful work instead of polished speculation.