Agentic Large Language Models
by Leiden University (LIACS): Max van Duijn, Michiel van der Meer, Aske Plaat, Niki van Stein
A free 472-page textbook on how LLM agents reason, act with tools and interact with each other, released into the public domain
Overview
Agentic Large Language Models is a 472-page textbook by Max van Duijn, Michiel van der Meer, Aske Plaat and Niki van Stein of LIACS, Leiden University, submitted to arXiv (2610.13455, cs.AI) on October 9, 2026 under a CC0 1.0 public domain dedication. It extends the authors' earlier survey of agentic LLMs, published in the Journal of Artificial Intelligence Research, and is the primary text for their Fall 2026 graduate course at Leiden. The book is organized around three pillars: reasoning, acting and interacting. Part I covers language modeling (tokenization, embeddings, attention, the training pipeline from pretraining through preference tuning, scaling laws and small distilled models). Part II covers reasoning models, chain-of-thought, self-reflection, reinforcement learning with verifiable rewards, vision-language-action and world models, and metalearning. Part III describes interaction styles and agentic memory. Part IV covers agentic retrieval (RAG, Self-RAG, multimodal retrieval), tool use through the ReAct loop and the Model Context Protocol, and applied assistants in shopping, medicine, finance, science and coding. Part V studies agent behavior, role-based multi-LLM orchestration and emergent behavior in multi-agent simulations. Optional chapters cover synthetic training data, mechanistic interpretability for agents and cognitive perspectives. Four case-study chapters pair with course assignments: building an LLM from scratch (the course uses Karpathy's nanochat), reasoning and reflection on the BALROG game benchmark, a multi-agent diagnosis framework, and simulating societies of LLM agents.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Book
- Cost
- Free
- Duration
- 472 pages; the companion Leiden course covers it in 14 weeks (6 EC, about 168 study hours)
- Provider
- Leiden University (LIACS): Max van Duijn, Michiel van der Meer, Aske Plaat, Niki van Stein
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How tokenization, embeddings, attention and transformer blocks combine into a working language model
- ✓The full LLM training pipeline from pretraining data through finetuning, preference tuning, inference and benchmarks
- ✓How scaling laws, small distilled models and efficiency trade-offs shape which model to build or deploy
- ✓How reasoning models use chain-of-thought, self-reflection and reinforcement learning with verifiable rewards
- ✓How agents retrieve knowledge with RAG, Self-RAG and agentic or multimodal retrieval pipelines
- ✓How agents call tools through the ReAct loop and connect to services with the Model Context Protocol
- ✓How role-based orchestration topologies coordinate several LLMs in one multi-agent system
- ✓How to evaluate reasoning and self-reflection on BALROG games and puzzles in a hands-on case study
Highlights
- •Free under CC0, so you can copy, adapt and redistribute it without asking permission
- •Covers the whole stack in one text, from transformer internals to MCP tool use and multi-agent simulation
- •Written as the textbook for a running Fall 2026 Leiden graduate course with a public syllabus and week-by-week schedule
- •Builds on the authors' JAIR survey of agentic LLMs, so the literature coverage is broad and referenced
- •Four case-study chapters tie the theory to coding assignments using nanochat, ReAct and the BALROG benchmark
- •Includes behavioral and multi-agent material (Theory of Mind, norm formation, collective behavior) that most engineering-focused agent books skip
Who It’s For
Best For
- ✓AI engineers who want a single structured reference on agent design instead of scattered blog posts
- ✓Graduate students or self-learners following a semester-length curriculum on agentic LLMs
- ✓Instructors looking for a free, openly licensed textbook to build an agents course around
- ✓Researchers interested in multi-agent behavior and simulated societies of LLM agents
Prerequisites
- •Python and PyTorch programming
- •Classical machine learning and deep learning basics (MLPs, CNNs, RNNs, backpropagation)
- •Reinforcement learning fundamentals such as Q-learning, DQN, PPO and SAC, per the course syllabus
- •NLP basics: tokenization, embeddings, attention and familiarity with Hugging Face
FAQ
What is Agentic Large Language Models?
Agentic Large Language Models is a free, CC0-licensed textbook from four Leiden University researchers, posted to arXiv on October 9, 2026. It is written for engineers and graduate students who know basic deep learning and want one structured text that runs from transformer foundations through reasoning models, tool use, RAG, MCP and multi-agent systems.
Is Agentic Large Language Models free?
Agentic Large Language Models is free to access.
What level is Agentic Large Language Models for?
Agentic Large Language Models is aimed at a intermediate audience. Recommended background: Python and PyTorch programming, Classical machine learning and deep learning basics (MLPs, CNNs, RNNs, backpropagation), Reinforcement learning fundamentals such as Q-learning, DQN, PPO and SAC, per the course syllabus, NLP basics: tokenization, embeddings, attention and familiarity with Hugging Face.
How long does Agentic Large Language Models take?
Expect roughly 472 pages; the companion Leiden course covers it in 14 weeks (6 EC, about 168 study hours). Most learners work through it at their own pace.
What will I learn from Agentic Large Language Models?
You'll learn: How tokenization, embeddings, attention and transformer blocks combine into a working language model; The full LLM training pipeline from pretraining data through finetuning, preference tuning, inference and benchmarks; How scaling laws, small distilled models and efficiency trade-offs shape which model to build or deploy; How reasoning models use chain-of-thought, self-reflection and reinforcement learning with verifiable rewards; How agents retrieve knowledge with RAG, Self-RAG and agentic or multimodal retrieval pipelines; How agents call tools through the ReAct loop and connect to services with the Model Context Protocol; How role-based orchestration topologies coordinate several LLMs in one multi-agent system; How to evaluate reasoning and self-reflection on BALROG games and puzzles in a hands-on case study.
Topics
Sources
This page was written from 4 sources, 3 on domains other than arxiv.org.