Skip to content

Start Here

If you just arrived at this book, start with one question: do you need an impressive demo agent, or a system that can survive production reality?

This book is written for the second case. It is most useful when read as one argument about how agent systems mature: from prompt-heavy prototypes to governed systems with trust boundaries, a policy layer, approvals, observability, evals, and lifecycle discipline.

Building agents is boring, but the result is staggering: discipline around trust boundaries, traces, approvals, and rollout turns a demo into a system that can be improved safely.

This page exists for one reason: to help you choose a reading route quickly.

If You Read Only One Thing

If you want the shortest entry into the book's thesis, read Chapter 1. Why an Agent Needs a Platform, Not Magic.

That chapter states the main claim plainly: a production agent system cannot be built as "a model plus some tools." It has to be designed as a governed operational system.

What Kind of Book This Is

This is not a guide to one framework and not a catalog of AI features. It is a practical architecture book for teams that need to run agents in real environments with write paths, human approvals, access boundaries, telemetry, evals, and explicit operational ownership.

A 30-Minute Route

If you have little time, read this path:

  1. Chapter 1. Why an Agent Needs a Platform, Not Magic
  2. Chapter 3. Security Perimeter and Trust Boundaries
  3. Chapter 8. Execution Model and Tool Catalog
  4. Part V. Reliability and Observability
  5. Chapter 18. Production Rollout Checklist

After that route, you should already have a working frame for:

  • where the real trust boundaries of an agent live;
  • what safe tool execution looks like;
  • why a smart model is not enough without traces, SLO, and evals;
  • what is required before the first serious rollout.

If you want to follow the running case

Follow the support-triage story: it starts with retrieval and safe tool execution, moves through duplicate-ticket recovery, traces, SLOs, and eval gates, then continues into rollout, ADLC, assurance, provenance, retirement, misalignment controls, telemetry, and registry. This is the best route if you want one incident-to-platform-contract path instead of abstract layers.

Reading Paths by Role

If You Are a Product Engineer

  1. Part I. Foundations
  2. Part II. Security Perimeter
  3. Part IV. Tools and Execution
  4. Part VII. Reference Implementation

This route is for moving quickly from an agent idea to a runnable architecture.

If You Are a Platform Engineer

  1. Chapter 2. Reference Architecture for a Safe Agent
  2. Part III. Memory and Knowledge
  3. Part IV. Tools and Execution
  4. Part V. Reliability and Observability
  5. Part VII. Reference Implementation

This route is for teams assembling a platform skeleton, not just a thin wrapper around one model.

If You Are a Security Engineer

  1. Part II. Security Perimeter
  2. Chapter 5. Why an Agent Needs Memory, and Why Memory Is Risky
  3. Chapter 9. Sandbox Execution and MCP as an Integration Contract
  4. Chapter 10. Idempotency, Retries, Rate Limits, and Rollback Boundaries
  5. Chapter 18. Production Rollout Checklist

This route is useful if you need to see not only model risk, but real execution risk.

If You Are a Lead or Architect

  1. Chapter 1. Why an Agent Needs a Platform, Not Magic
  2. Part V. Reliability and Observability
  3. Part VI. Organizational Model
  4. Chapter 18. Production Rollout Checklist

This route is for keeping an initiative inside real operational discipline instead of shipping only a demo.

If You Want Code and Artifacts First

If executable support matters more than linear reading, start here:

This is useful if you want a runtime skeleton, policy contracts, memory paths, telemetry, and rollout artifacts right away.

If You Need To Solve One Specific Problem Fast

Safe Tool Execution

Memory and Retrieval

Observability, Evals, and Rollout

What To Keep Open Next to the Book

If this book feels closer to you than another AI landing page about "autonomy," you are in the right place.