Designing AI Agents: Principles, patterns, and best practices

Get the eBook free when you register your print book at Manning.

AI agents promise to automate work on an unprecedented scale—even for tasks requiring reasoning and complex multi-step processes. But where do you start? How do you budget your dev time and token spend? What do you do with an agent that “almost works?” How do you scale or improve a multi-agent system? This book answers all these questions and more. In this enlightening book, author Jia Huang, a senior AI researcher at the Singapore-based Agency for Science, Technology and Research (A*STAR), presents an innovative two-axis framework blending seven cognitive functions—perception, memory, reasoning, action, reflection, collaboration, and governance—with six topologies—chain, route, parallel, orchestrate, hierarchy, and loop.

In this book, you’ll learn how to establish agent architectures that manage costs and take governance seriously from day one. This innovative book explores 27 reusable patterns that you can apply to your own agentic systems confidently. Each pattern has been stress-tested at scale, with over 10,000 engineers applying them to ship production agents in banks, manufacturers, and AI startups. You’ll appreciate how this book guides you toward system and harness design that imposes certainty and reliability on the non-deterministic behavior of LLM-driven agents. Once you catch author Jia Huang’s vision, you’ll stop asking “which tools?” and start asking "which patterns?".

Unlike other “agentic patterns” books that dwell on abstract theory, every chapter in this practical guide grows a single running example. You’ll incrementally build Argus, a code-review agent that evolves from a 50-line reasoning-and-action loop all the way to a production-grade system. As you steadily upgrade Argus, you’ll learn both which patterns to use and when and why to apply them. Plus, full case studies exploring agents working as a DevOps incident responder, compliance reviewer, and research synthesis agent show the methodology in action across diverse domains. And, as with all Manning books, you’ll find a clearly defined learning path, thoughtfully edited and readable text, and our promise that everything is accurate and reliable.

What's inside

• A systematic framework with 27 design patterns
• Agent governance as architecture
• A shared vocabulary for discussing agents
• Reference harnesses, including Claude Code, Cursor, and OpenClaw

About the reader

For engineers who know the basics of agent design and want to deliver reliable, cost-effective agentic AI.

About the author

Jia Huang is an AI researcher at A*STAR Singapore. He spent a decade as a senior SAP consultant at Accenture, has authored several bestselling Chinese-language AI and agent design books, and teaches agent engineering to 10,000+ engineers through his GeekTime course.