Cover of Foundations of Agentic AI Systems, Volume I: Models, Learning, and Decisions, from First Principles, by Fatih Nayebi, on a navy ground

Volume I · October 2026

Foundations of Agentic AI Systems, Volume I

Models, Learning, and Decisions, from First Principles

First principles for production agentic systems that can be judged.

A model can pass every offline evaluation and still break in production, because nobody designed the controls around it: a policy nobody versioned, a tool that trusted its input, an action no one could take back. This book is about agents that act, and about what an agent must show before it is allowed to.

Hardcover on Amazon Companion code

About the book

Volume I follows one running example, a refund agent, from its first day, when it only drafts decisions for a person, to the day every proposal it makes stands beside its evidence, its baseline and a verdict. It still pays nothing. Each method arrives as something the agent must show before it may act: calibrated probabilities and the costs behind a threshold, causal effects, neural networks and language models, reinforcement learning and off-policy evaluation.

The derivations are worked, most results come with a number you can check, and the examples run in a companion code repository whose stored outputs the book prints. Often the evidence says no. A learned dispatcher that reads a 91.7% saving off its own estimates costs more than three times the plain rule on shifts it never saw, and the book shows why holding it back is the right call.

Volume II, Agents, Operations, and Governance, in Production, takes the same agent to its launch gate.

Contents

Part I · Orientation and Agentic Mindset

  1. 1 What Is Agentic AI?
  2. 2 Architecture, Components and Contracts
  3. 3 Decision Strategy and Evidence Requirements

Part II · ML Foundations

  1. 4 ML Data and Evaluation
  2. 5 Supervised Learning and Calibrated Risk
  3. 6 Semi- and Self-Supervised Learning and Unlabeled Structure
  4. 7 Causal Inference and Experiments
  5. 8 Delegated Model Building

Part III · Deep Learning Foundations

  1. 9 Neural Networks and Backpropagation
  2. 10 Training Optimization and Regularization
  3. 11 Autoencoders and NLP Representations
  4. 12 Sequence Processing, Transformers and Foundation Models
  5. 13 Generative Models and Simulation
  6. 14 Prompting, Fine-Tuning and Retrieval Adaptation

Part IV · RL and Decision-Making

  1. 15 Constrained Optimization for Agentic Decisions
  2. 16 Reinforcement Learning Foundations
  3. 17 Exploration, Bandits and Exposure Budgets
  4. 18 Deep Reinforcement Learning
  5. 19 RLHF and Preference Alignment
  6. 20 Typed Decision Models and Calibrated Decisions
  7. 21 Practical RL and Off-Policy Evaluation
Fatih Nayebi

About the author

Fatih Nayebi

Fatih Nayebi is Assistant Professor in Information Systems at McGill University, where since 2019 he has taught the five graduate courses these volumes hold: Enterprise Data Science, Machine Learning in Production, Introduction to AI and Deep Learning, Applications and Architectures of Deep Learning, and Designing and Building Agentic AI Systems. As founder and CEO of OODARIS AI, he builds real-world, production multi-agent systems that make autonomous decisions, combining machine learning, ML engineering, deep reinforcement learning, operations research and software engineering, and that act only within the authority they have earned.

More about Fatih OODARIS AI

Also in the series

All books

Autonomy should rise only as evidence rises. · Published by Gradient Divergence, Montréal.