CHAPTER 15 · Deep Research Agents and How They Are Trained · 1 / 10
The architecture
The "Deep Research of Deep Research" survey describes a clean architecture, explicitly inspired by Anthropic's research agent. When a query arrives, the system confirms intent, then a MasterAgent enters a research loop: it thinks through the approach and saves its plan to memory, spawns specialized SubAgents with specific subtasks, each of which searches and evaluates results in its own context and returns findings, and the MasterAgent synthesizes and decides whether more research is needed. When enough is gathered, a ReviewAgent checks that every claim is properly attributed to a source before the cited report goes back to the user. That is the orchestrator pattern (Chapter 11) plus the verifier (the ReviewAgent), specialized for research.