Agentic Large Language Models, a Survey
by Leiden University (Plaat, van Duijn, van Stein, Preuss, van der Putten, Batenburg)
A peer-reviewed map of agentic LLMs organised as reason, act and interact, with a five-point research agenda.
Overview
Agentic Large Language Models, a Survey is written by six Leiden University researchers: Aske Plaat, Max van Duijn, Niki van Stein, Mike Preuss, Peter van der Putten and Kees Joost Batenburg. It first appeared on arXiv in March 2025, was revised to v3 on 22 November 2025, and was published in the Journal of Artificial Intelligence Research (vol. 84) on 30 December 2025 under a CC BY-SA 4.0 licence. It defines agentic LLMs as models that reason, act and interact, and gives each category a section. Section 2, Reasoning, covers multi-step chain of thought, interpreters and debuggers, tree search, self-reflection and prompt improvement, and retrieval augmentation. Section 3, Acting, covers world models, vision-language-action models, robot planning, action tools, computer and browser tools, agent interoperability, safety and security, and domain assistants for medicine, science and trading. Section 4, Interacting, covers conversation, strategic behaviour, theory of mind, role-based multi-agent teamwork, simulated societies and emergent social norms. Each section ends with an in-depth discussion: CoT and self-reflection, the AI Scientist, and CAMEL and Generative Agents. It closes with a research agenda on training data from inference-time behaviour, hallucination and stable behaviour, agents at scale, self-reflection, and safety. Three of the authors now teach Leiden's Fall 2026 Agentic LLMs course from a textbook built on this framework.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Paper
- Cost
- Free
- Duration
- Long-form journal survey; several hours for a full read, self-paced
- Provider
- Leiden University (Plaat, van Duijn, van Stein, Preuss, van der Putten, Batenburg)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Classify agent techniques into the reason, act and interact taxonomy used throughout
- ✓Compare multi-step reasoning methods: chain of thought, interpreters and debuggers, and tree search
- ✓Understand how self-reflection and prompt-improvement loops raise LLM decision quality
- ✓See how retrieval augmentation connects reasoning to tool use in agent systems
- ✓Survey action models including world models and vision-language-action models for robotics
- ✓Weigh computer, browser and interoperability tools against their safety and security risks
- ✓Analyse role-based multi-agent collaboration, simulated societies and emergent social norms
- ✓Identify open problems: hallucination, stability at scale, self-reflection, safety and training data
Highlights
- •Peer-reviewed in JAIR (vol. 84, December 2025), not only an arXiv preprint
- •Its three-way taxonomy shows how the categories reinforce each other, e.g. retrieval enables tool use and reflection improves multi-agent collaboration
- •In-depth case discussions of the AI Scientist, CAMEL and Generative Agents ground the theory
- •Serves as the backbone of Leiden University's Fall 2026 Agentic LLMs master's course
- •Openly licensed (CC BY-SA 4.0), with an HTML version and a companion site
Who It’s For
Best For
- ✓AI engineers who want a structured mental model of agent design before choosing frameworks
- ✓Researchers and graduate students starting a literature review on LLM agents
- ✓Technical leads assessing where agentic systems are mature and where they remain risky
- ✓Instructors building an agents curriculum around reasoning, acting and interaction
Prerequisites
- •Working knowledge of how LLMs are trained and prompted (pretraining, fine-tuning, chain of thought)
- •Comfort reading academic ML papers and following citations
- •No coding required; it is a literature survey, not a hands-on tutorial
FAQ
What is Agentic Large Language Models, a Survey?
A peer-reviewed survey, published in the Journal of Artificial Intelligence Research in December 2025, that organises the agentic-LLM literature into three categories: reasoning, acting and interacting. It is for engineers and researchers who want a structured mental model of agent design choices before choosing frameworks or reading individual papers.
Is Agentic Large Language Models, a Survey free?
Agentic Large Language Models, a Survey is free to access.
What level is Agentic Large Language Models, a Survey for?
Agentic Large Language Models, a Survey is aimed at a intermediate audience. Recommended background: Working knowledge of how LLMs are trained and prompted (pretraining, fine-tuning, chain of thought), Comfort reading academic ML papers and following citations, No coding required; it is a literature survey, not a hands-on tutorial.
How long does Agentic Large Language Models, a Survey take?
Expect roughly Long-form journal survey; several hours for a full read, self-paced. Most learners work through it at their own pace.
What will I learn from Agentic Large Language Models, a Survey?
You'll learn: Classify agent techniques into the reason, act and interact taxonomy used throughout; Compare multi-step reasoning methods: chain of thought, interpreters and debuggers, and tree search; Understand how self-reflection and prompt-improvement loops raise LLM decision quality; See how retrieval augmentation connects reasoning to tool use in agent systems; Survey action models including world models and vision-language-action models for robotics; Weigh computer, browser and interoperability tools against their safety and security risks; Analyse role-based multi-agent collaboration, simulated societies and emergent social norms; Identify open problems: hallucination, stability at scale, self-reflection, safety and training data.
Topics
Sources
This page was written from 5 sources, 3 on domains other than arxiv.org.