A taxonomy worth knowing

The DR survey offers a taxonomy that is genuinely useful for any agent designer, not just research ones.

Static vs dynamic workflows. A static workflow runs a predefined pipeline (ideate, then experiment, then report). It is easy to build and predictable, but each new task needs a hand-tailored pipeline, so it generalizes poorly. A dynamic workflow lets the agent reconfigure its plan on the fly based on what it learns. More flexible, more general, more demanding on the model. This is the same "workflow vs agent" tradeoff Anthropic's "building effective agents" guidance raises: use a fixed workflow when the steps are known, a full agent loop when exploration is needed.

Planning strategies (how the agent involves the user before acting):

Strategy	Behavior	Example
Planning-only	Plan straight from the prompt, no clarification	Grok, Manus
Intent-to-planning	Ask clarifying questions first, then plan	OpenAI DR
Unified intent-planning	Draft a plan, then ask the user to confirm or edit it	Gemini DR

Single-agent vs multi-agent. A single agent folds planning, tool use, and execution into one cognitive loop (the ReAct pattern: reason, act, observe). It is coherent and, crucially, easy to optimize end to end with reinforcement learning. A multi-agent system splits work across specialists with a coordinator, which scales to parallel, complex tasks but is much harder to train end to end. That training difficulty is a recurring theme: multi-agent systems are powerful but hard to optimize as a whole.

Memory mechanisms for long research (the same three moves from Chapter 5, confirmed across the literature): extend the context window (Gemini's million tokens), compress intermediate steps (summarize as you go, like Search-o1's "Reason-in-Documents"), or use external structured storage (files, vector databases, even knowledge graphs). Manus's file memory from Chapter 14 is the external-storage option.