Agentic Engineering

From neural networks to dependable agent systems

An executable, systems-first textbook for senior generative-AI and agent engineers.

Agentic Engineering

Modern agents are not a prompt wrapped in a loop. They are probabilistic decision-makers embedded inside ordinary software systems: APIs, stores, queues, policies, sandboxes, evaluators, deployment pipelines, and human organizations. Engineering them well means understanding both the model and everything that constrains what the model can do.

This book builds that understanding from one starting point: you know neural networks, backpropagation, and basic Python, and we assume nothing else. We derive the language-model substrate, follow one request through inference and serving, build retrieval and agent mechanisms from scratch, and then put those mechanisms under evaluation, security, reliability, and operational pressure. The habit we keep throughout is simple.

Add autonomy only when measured value pays for its reliability, cost, security, and operator burden.

flowchart LR
    M[Model substrate] --> P[Programmed model calls]
    P --> R[Retrieval and context]
    R --> A[Bounded agent]
    A --> E[Evaluation and security]
    E --> O[Production operations]
    O --> C[Defended capstone]

Figure 1: The book grows one dependable system outward from the model.

How this book teaches

We teach the way Dive into Deep Learning teaches: you watch working code produce real output, narrated by a knowledgeable “we”. Every mechanism is developed inline in executable cells whose printed values, generated text, and plots appear right beneath them, so you can reproduce every number yourself. Figures are there to build intuition, not to decorate. Every build is deterministic and runs offline by default.

There is one linear path, Chapters 1 through 32, and a reader already fluent in transformers and serving can start at Chapter 12. Part IX (Chapters 33–40) is the agentic system-design interview: one reusable seven-move framework, then seven worked designs, each teaching a whole problem family. Appendix A is a just-in-time bridge for the occasional mathematical or systems background a chapter leans on; Appendix B holds the evidence templates, sizing worksheets, and drills; Appendix C holds the volatile facts, so the chapters can stay about mechanisms.

Begin with Chapter 1, or take the book with you: download the EPUB.