Preface

The demo is not the system

Almost anyone can wire an LLM to a few tools in an afternoon and watch it accomplish something that looks like magic. Turning that loop into infrastructure that thousands of agents can run on — concurrently, durably, observably, under budget, and without leaking your data or your customers' — is a different discipline entirely. That gap is the subject of this book.

A modern agentic platform is not a framework. It is a layered system in which a probabilistic, slow, expensive, and occasionally adversarial component — the language model — is harnessed inside deterministic machinery that makes its behavior safe to depend on. The framework you pick (LangGraph, the OpenAI Agents SDK, CrewAI, Letta, or one you write yourself) occupies exactly one layer of perhaps twelve. The other eleven are what separate a clever notebook from a platform.

This book builds the platform from the bottom up. We begin by deriving what an agent fundamentally is and what it fundamentally needs, with no reference to any product. From those needs, each subsequent layer falls out by necessity: because LLM calls fail and runs are long, we need durable execution; because calls are expensive and multi-provider, we need a gateway; because agents act in the world, we need tools, sandboxes, and identity; because they are non-deterministic, we need observability and evaluation. Nothing in the architecture is arbitrary. Every box on the master diagram earns its place by solving a specific, nameable problem.

You should read this if you are building, buying, or operating agent infrastructure and you want to understand the why beneath the what. By the end you will be able to draw the full stack from memory, justify each layer, choose components against your constraints, and recognize the failure modes before they reach production.