Preface

An agent does not achieve production reliability simply by appending more instructions to its prompt. The chasm between an impressive demo and a system capable of handling real-world user traffic lies in less glamorous architectural decisions: what actions it is permitted to take, what it is forbidden from doing, who authorizes side effects, and how its execution trace is verified.

This volume compiles those architectural decisions. It starts from the formal definition of an autonomous agent and its sharp boundary with deterministic workflows, advancing toward the minimum viable architecture required to run agents reliably: gateways, control harnesses, context boundaries, long-term memory, and typed tools. The included code examples do not form a locked, opinionated framework. Instead, they illustrate a transparent structure that can be understood, adapted to the real-world constraints of your product, and translated into deterministic test assertions. The code is meant to be read, debated, and refactored.

This work deliberately distinguishes between principles, implementations, and tools. Principles guide architectural design; implementations are perishable design choices; tools belong to transient market moments. Maintaining that distinction is vital: it enables you to evolve your architecture without turning every vendor change into an organizational crisis.

Each chapter features worked exercises designed to dissect trade-offs, not merely showcase happy paths. You may read this volume linearly or consult specific chapters targeting concrete operational challenges. The critical reading workbook, glossary, and quick reference index provide entry points for the latter approach.

This first volume does not claim to solve all of agentic systems engineering. It articulates the hard questions engineering teams must answer before declaring a system production-ready, establishing the shared vocabulary required to do so.

Sergio Perez Ruiz
Lima, 2026