This is a book I wanted to like more than I do, and a book that I think is important whether I like it or not. Enterprise Architecture in the Age of Agentic AI is the first sustained attempt I have seen to re-frame the EA discipline for a world in which agents — autonomous, goal-seeking, tool-using software systems — are first-class participants in business processes rather than oddities at the edge of the architecture diagram.

The book is uneven. Some sections are genuinely original; some are restatements of existing EA orthodoxy with “agent” substituted for “service.” But the project of the book is the right project, and we need more books like it. Dahdour got there early.

The thesis

The thesis, as I read it: agentic AI breaks several load-bearing assumptions of traditional EA. Specifically:

  • The assumption of deterministic interfaces. Traditional EA assumes that the contract between two components is a stable thing you can document. Agentic systems have contracts that are partly statistical and that drift with model updates.
  • The assumption of clear ownership. Traditional EA assumes that each capability has an owner who can answer questions about it. Agentic systems often have emergent capabilities that no team has explicitly chosen to build.
  • The assumption that governance can be perimeter-based. Most EA governance assumes that you can put a wrapper around a system and inspect the inputs and outputs. Agentic systems compose at runtime in ways that make perimeter governance partial at best.

Dahdour argues that the response to this is not to give up on EA but to raise its level of abstraction — to govern the constraints and invariants that agents must respect, rather than the specific interfaces they will use. This is the right framing. It also rhymes strongly with the fitness-function framing in Building Evolutionary Architectures, which the book references but does not lean on as hard as I would have.

What the book gets right

The capability-versus-agent distinction. Dahdour argues that business capabilities are still the right primary unit of EA, and that agents are implementations of capabilities (or parts of them). This keeps the EA discipline grounded in the business view rather than chasing whatever the current generation of models can do. I think this framing will age well.

The treatment of the “agent registry.” The book argues that organisations will need a registry of agents analogous to (but distinct from) the service catalog. The registry needs to track not just what each agent does but what each agent is allowed to do, what data it can see, and what humans are accountable for its behaviour. This is correct, it is non-obvious, and it is the kind of practical recommendation EA practitioners can actually act on.

The honesty about risk. Dahdour does not pretend that the existing risk frameworks transfer cleanly. The discussion of compounded autonomy risk — the way risks combine non-linearly when agents call other agents — is the best treatment of this I have read in any EA book.

What the book gets less right

The case studies are thin. The book is too new for the case studies to be load-bearing. The examples are mostly hypothetical, and where they are not, they are the same five examples (customer-service agents, code-review agents, agentic data pipelines) that every other 2026 book is also using. This is not Dahdour’s fault, but it is a limitation.

The technical depth is uneven. Some sections are precise about agent architecture (the chapter on tool-use semantics is good); others hand-wave through “the LLM decides.” A reader who has built agentic systems will notice the variance.

The book is too long. It could be a hundred pages shorter without losing anything important. The first three chapters and the last two are the load-bearing parts; the middle chapters could be skimmed.

How it sits next to existing EA

The book is in conversation with the existing canon in a way I appreciated. Enterprise Architecture As Strategy is invoked explicitly, and the operating-model quadrant survives the agentic-AI re-framing with surprisingly little adjustment. The Foundation for Execution concept extends naturally to “Foundation for Autonomous Execution,” which is a clunky phrase but the right idea.

The book is in less conversation with the platform-engineering and team-topology literature than I would have liked. The implications of agentic systems for team design — who owns the agent? who is on call for it? what does an “incident review” for an agent’s bad decision look like? — are mostly ducked. There is room for another book on exactly that, and I hope someone writes it soon.

Reading note

The book is Kindle-first; the print edition (if it exists in your region) is print-on-demand and the diagrams suffer. The Kindle edition is the better read. The book is dense enough that I would not recommend an audiobook even if one exists.

It pairs naturally with Agentic Architectural Patterns, which covers the pattern-level material Dahdour mostly skips, and with Enterprise Architecture As Strategy, which provides the operating-model grounding Dahdour assumes.

Who I’d give this to

  • A head of EA who has been asked by the board what the AI strategy means for the architecture function.
  • A CTO trying to put guardrails around an internal agentic platform before it ships to customers.
  • A principal engineer who is moving from traditional service architecture into agentic-system architecture and needs the EA vocabulary that the platform-engineering literature is missing.
  • A senior risk or compliance leader trying to engage substantively with AI governance.

I would not give it to:

  • A junior engineer. The book is at the wrong altitude.
  • A reader who wants implementation patterns. For that, go to Agentic Architectural Patterns instead.

What I think the book is doing for the field

The most important thing this book does is make the conversation respectable. Before this book, “enterprise architecture for agentic AI” was a topic you would raise carefully in a senior-architect Slack channel, get four useful replies and twelve hot takes, and put aside for another quarter. This book is the artefact that lets you walk into an executive conversation with a coherent position. Even where the book is wrong, it is wrong in legible ways that can be argued with. That is most of what you want from a foundational text.

Bottom line

The book is uneven but necessary. Read it. Argue with it. Read it again in a year. The 2027 second edition (if it exists) will be the better book, because the field will have moved, and Dahdour will be the person best placed to write it. In the meantime, this is the one to read.

For the patterns-level companion, go to Agentic Architectural Patterns. For the operating-model grounding it assumes, go to Enterprise Architecture As Strategy.