Book¶
This is the main entry page for the book itself. If you want the shortest route into the manuscript, start with Part I. Foundations. If you want the structure and publication status first, open the Book Plan.
What This Book Promises¶
This book argues one core thesis: agents need a platform, not magic.
Building agents is boring, but the result is staggering: instead of a flashy one-off trick, you get a system that can be constrained, observed, shipped, and improved without guessing.
By reading it, you should be able to:
- decide when an agent is justified and when a normal workflow is enough;
- understand the minimum platform layers required before risky action is allowed;
- follow one governed run through policy, execution, evidence, approval, rollout, and lifecycle control;
- reason about memory, evals, provenance, retirement, and operator accountability as one operating model.
Running support case
One way to read the book is to follow the support-triage case: from retrieval and tool execution to duplicate-ticket recovery, traces, SLOs, eval gates, ownership, reference runtime, policy, rollout, ADLC, assurance, provenance, retirement, misalignment controls, telemetry, and registry. That turns the chapters from a set of topics into one reviewable story about how an incident becomes a platform contract.
Recommended Reading Path¶
If you want the shortest useful route, read in this order:
- Part I. Foundations
- Part II. Security Perimeter
- Part III. Memory and Knowledge
- Part IV. Tools and Execution
- Part V. Reliability and Observability
- Part VI. Organizational Model
- Part VII. Reference Implementation
- Part VIII. Agent System Lifecycle
Stability Shortcut¶
The book has two practical layers:
Stable core: Parts I-VII, especially Chapters 1-12 and 18;Fast-moving layer: Chapter 13, Part VIII, and research-heavy appendix pages.
If you are reading for the first time, start with the stable core and return to the faster layer after that.