Agent Design Patterns (Elements, components, and architecture)

AI agents are showing up everywhere—and most of them are being built by trial and error. This book distills the experience of a growing community of agent builders into 20+ composable, reusable patterns for building AI agents that are reliable, efficient, controllable, and easy to reason about. In this uniquely valuable book, author and NVIDIA Research scientist Peter Belcak gives you a durable engineering vocabulary and design intuition that outlasts any new model release.

Each pattern included here is production tested and designed to click together as you construct transparent, testable, deployment-ready agentic applications. You’ll appreciate the familiar presentation style, with a simple template of a named problem, a clear solution, benefits, drawbacks, popular variants, and rules for composition with other patterns. Informative diagrams, worked examples, and Belcak’s clear mathematician-turned-engineer voice make even the most subtle patterns easy to internalize.

By the end, you’ll be able to read an underperforming agent, pinpoint the exact quality that is lacking, and reach for the specific pattern that addresses it. You’ll lower per-run costs without redesigning your agent, and even build self-improving agents that auto-tune their own prompts.

What's inside

• 20+ composable, framework-agnostic agent design patterns
• A visual language for designing and communicating agent architectures
• Techniques to boost reliability, cut cost, and score confidence
• Human-in-the-loop patterns for safe, controllable agents
• Self-improving agents that auto-tune their own prompts

About the reader

For software and AI/ML engineers, solutions architects, and applied AI practitioners building production-grade agents.

About the author

Peter Belcak is an AI Research Scientist, currently at NVIDIA Research. Prior to that he was at Meta GenAI and ETH Zurich. He has contributed across the entire AI stack, from AI efficiency, through efficient model training and tuning frameworks to NVIDIA’s Enterprise AI agent frameworks and products. He currently researches methods for improving the reliability of agentic systems and the formation of logically correct reasoning systems, and leads product efforts on LLM hallucination elimination and LLM advertising.