The Agentic AI Builder's Handbook
Principles, Patterns, Scaffolds, and Harnesses
A practical handbook for building agentic AI applications you can control, evaluate, and ship.
"In theory there is no difference between theory and practice, while in practice there is." — Attributed to various; popularized by Yogi Berra
Why this book (and why now)
There are plenty of social media posts, demos, and blog articles that show how to build an agent quickly. They help you get started. But they fall short where it matters most in real systems.
How do you constrain agent behaviour so it stays predictable? How do you evaluate it beyond toy examples? How do you make it safe, auditable, and operable in production? How do you keep it working as models, policies, data, and tools change?
This handbook answers these questions. It covers the engineering discipline required to move from "it works in a demo" to "it works in the enterprise."
What this book is
This handbook is for builders moving beyond prompt demos into production-grade agentic systems. These are systems that can plan, use tools, operate within policy boundaries, and be measured and improved over time.
The handbook is organised around four ideas:
- Principles — how to make good design choices
- Patterns — reusable solutions to recurring problems
- Scaffolds — the control structures that shape agent behaviour (inner loop)
- Harnesses — the environments that test, measure, and monitor agents (outer loop)
Running examples
This handbook uses three practical examples that build progressively across chapters.
The HR Leave Agent is the primary example. It appears in most chapters and demonstrates the full lifecycle from design to production. The IT Help Desk Agent illustrates multi-turn conversation patterns. The Invoice/Procurement Assistant shows tool orchestration scenarios.
Chapter 14 provides complete implementations of these examples.
What this book is not
This handbook does not teach LLM fundamentals, prompt engineering basics, or machine learning theory. It assumes you can already call a model API and get useful output.
It is not a framework tutorial. It does not walk through LangChain, CrewAI, AutoGen, or any specific SDK. The patterns here are framework-agnostic and apply regardless of which libraries you choose.
It is not an academic survey of agent research. It cites research where it informs engineering decisions, but the focus is on building and shipping, not on benchmarks or theoretical architectures.
It is not a global regulatory compliance guide. Governance examples focus on Singapore (MAS, PDPA, AI Verify) as a concrete illustration. The frameworks are adaptable, but you will need to map them to your own jurisdiction.
It is not a cookbook of 50 agent recipes. It uses one primary running example (HR Leave Agent) applied end-to-end across all chapters. Depth on one realistic system teaches more than breadth across toy examples.
Who this book is for
This book is written for people building agentic AI in real products and workflows:
- Software engineers and AI engineers shipping LLM/agent features
- Tech leads and architects designing agentic systems and platforms
- Product builders who need reliable behaviour, clear UX controls, and measurable outcomes
- Domain SMEs and ops owners partnering with engineering to define "what good looks like"
- Risk, governance, and assurance teams who need evidence, auditability, and lifecycle controls
If you want the fastest way to demo an agent, parts of this book may feel heavier than you need. If you want to build something you can deploy, monitor, and defend, this is for you.
Chapter overview
| Part | Chapters | Focus |
|---|---|---|
| Foundation | 1–3 | Why, What, Principles |
| Design | 4–5 | Scaffolds & Harnesses |
| Measurement & Optimization | 6–9 | Evaluation, Optimization & Patterns |
| UX & Operations | 10–11 | UX & Production |
| Governance | 12–13 | Governance & Teams |
| Examples | 14 | Case Studies |
| Reference | Appendices | Templates & Checklists |
How to read this handbook
You do not need to read this handbook cover to cover. Choose a path based on your experience level or role.
By experience level:
| If you are... | Start with... | Then... |
|---|---|---|
| New to agents | Ch 1–3 (Foundation) | Ch 4–5 (Design) |
| Building now | Ch 4–7 (Design to Evaluation) | Ch 8–9 (Optimize → Patterns) |
| Shipping soon | Ch 8–12 (Optimize → Prod → Gov) | Ch 14 (Case Studies) |
| Debugging issues | §14.6 (Failure Cases) | §11.4 (Observability) |
By role (with estimated reading time):
- Product Managers — Ch 1, 3, 10, 11, 14 (~3 hours)
- Software Engineers — Ch 2, 4, 5, 9, 11 (~5 hours)
- Architects — Ch 2, 4, 9, 11, 12 (~4 hours)
- Governance and Risk — Ch 1, 3, 6, 12, 14 (~4 hours)
- ML/AI Engineers — Ch 2, 6, 7, 8, 9, 11 (~5 hours)
- Engineering Managers — Ch 1, 3, 11, 13, 14 (~4 hours)
For detailed guidance by role and task, see the Reading Paths guide.
Start here
Begin with the Table of Contents for a complete chapter listing. Then read Chapter 1 — Why Agentic AI, Why Now for strategic context.
About
This book grew out of the countless energetic discussions and explorations I shared with my colleagues - Josey Mathew, Leng Yiren, Dimitri Bintein, Ying Shaowei, Hasala Marakkalage, Joe Jimeno, Edward Widjaja, and Ritchie Ng.
Version: 1.1 Last updated: February 2026