Many real-world questions cannot be answered responsibly from one model call. They require searching across sources, reading the original material, separating evidence from interpretation, comparing conflicting claims, and writing an answer the 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 cited report. Its job is not to sound confident. Its job is to make the path from question to evidence visible.

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