CH 15 · Deep Research Agents and How They Are Trained · 01/11
Read as one pageDeep Research Agents and How They Are Trained
A deep research agent plans, spawns subagents to search, and checks claims against sources. It is trained with reinforcement learning, not just prompted.
11 slides · 7 min read
In this chapter
- 01The architecture
- 02A taxonomy worth knowing
- 03Fact-checking: the research version of "do not trust yourself"
- 04How they are trained
- 05Levels of automation and where agents still fall short
- 06Why it matters
- 07How it works: a research loop with planning, retrieval, and a review pass
- 08Code MVP: a deep research loop
- 09Connecting to the bigger picture
- 10Key takeaways
A deep research (DR) agent is an agent specialized for open-ended information work: you ask a complex question, and it browses, retrieves, reasons across many sources, and writes a structured, cited report. OpenAI's Deep Research, Gemini Deep Research, Perplexity, and Grok DeepSearch are the famous examples. They are built on the exact harness ideas from this guide (loop, tools, planning, memory, multi-agent), so studying them sharpens those ideas, and they add a new dimension we have mostly skipped so far: how you train an agent to be good, not just prompt it. This chapter draws on two academic surveys: "Deep Research Agents: A Systematic Examination and Roadmap" and "Deep Research of Deep Research: from transformer to agent."
The formal definition from the survey: DR agents are "AI agents powered by LLMs, integrating dynamic reasoning, adaptive planning, and iterative tool use to acquire, aggregate, and analyse external information, culminating in comprehensive outputs for open-ended informational research tasks." Compared with plain RAG (which boosts factual accuracy but does not sustain reasoning) and plain tool use (which depends on fixed workflows), DR agents add autonomy, deep reasoning, dynamic planning, and real-time adaptation.
Finished this idea?
Your learning progress stays in this browser.