1. Why These Three Books
Most lists of AI engineering books are ten titles long and tell you to read all of them. This one is three, chosen because each answers a different question, and because together they cover the job people now call AI architect: designing systems where some of the components are models.
| Book | Authors | Published | The question it answers |
|---|---|---|---|
| Fundamentals of Software Architecture, 2nd edition | Mark Richards, Neal Ford | March 2025 | How to think like an architect: characteristics, modularity, styles, decisions |
| Software Architecture: The Hard Parts | Neal Ford, Mark Richards, Pramod Sadalage, Zhamak Dehghani | October 2021 | What breaks when a system is distributed: coupling, data, transactions, contracts |
| AI Engineering | Chip Huyen | December 2024 | What changes when a component is a foundation model: evals, RAG, agents, inference |
The first two are from the same family of authors and were written to be read in that order. The third comes from a different community. That gap is the interesting part: where they agree, they usually don't know they agree.
Links to the publisher's pages: Fundamentals of Software Architecture, Software Architecture: The Hard Parts, AI Engineering. Chapter names below are from their published tables of contents.
2. The Idea in the Middle: An Eval Is a Fitness Function
Start with the overlap that matters most. In the architecture books, a fitness function is an automated check that tells you whether a system still has a property you care about. Chapter 6 of Fundamentals is called "Measuring and Governing Architecture Characteristics", and The Hard Parts attaches a fitness function to every decomposition pattern it describes.
In AI Engineering, an eval is an automated check that tells you whether a model-backed feature still does what you need. Huyen spends two chapters on it, covering exact evaluation, AI as a judge, and how to design an evaluation pipeline.
One community writes these checks against coupling and latency. The other writes them against faithfulness and tone. They are the same tool, used for the same reason: you can't keep a property you don't test for. An architect who already runs fitness functions in CI knows how to run evals in CI. An AI engineer who already gates releases on an eval set is already doing architecture governance, and can extend the habit to cost, latency and coupling.
Three more ideas sit in the centre with it.
| Shared idea | In Fundamentals | In The Hard Parts | In AI Engineering |
|---|---|---|---|
| Every choice is a trade-off | Stated as a law of software architecture, and the reason for the phrase "least worst architecture" | The closing chapter is called "Build Your Own Trade-Off Analysis" | Build versus buy, RAG versus finetuning, cost versus latency versus quality |
| Fitness functions ≈ evals | "Governance and Fitness Functions" for measuring architecture characteristics | Fitness functions attached to each decomposition pattern | Two chapters on evaluation, including AI as a judge and designing an evaluation pipeline |
| Orchestrated workflows | Mediated event-driven architecture and orchestration-driven styles | Orchestration versus choreography, and eight transactional saga patterns | Agent planning and "AI pipeline orchestration" |
| Schemas as contracts | Contracts between a microkernel core and its plug-ins | Strict versus loose contracts, and stamp coupling | Structured outputs from a model |
The workflow row deserves a second look. A saga is a long transaction split across services, with a plan for what to do when a step fails. An agent run is a long task split across model and tool calls, with the same need. The eight saga patterns are a ready-made vocabulary for multi-agent orchestration: who coordinates, whether steps wait for each other, and what "done" means when something fails halfway.
3. What Two Books Share: The Ideas That Transfer
These are the overlaps between pairs. Each is an architecture idea with a direct use in an AI system, or the reverse.
| Shared by | Idea | What it means for an AI system |
|---|---|---|
| Both architecture books | ADRs | A written record of a decision and why. Use them for model choice, RAG versus finetuning, and vendor lock-in |
| Both architecture books | Architecture quantum; static and dynamic coupling | The smallest part you can deploy on its own. A feature and the model it calls are coupled at runtime, so they fail together |
| Both architecture books | Stamp coupling | Passing a whole record when one field is needed. With an LLM this is paid for per token |
| Fundamentals + AI Engineering | Observability | A request is a tree of model and tool calls, so it has to be traced, not just timed |
| Fundamentals + AI Engineering | Latency and cost as characteristics | Treat them like any other "-ility": name them, measure them, pick a few to protect |
| Fundamentals + AI Engineering | Layered thinking | Huyen's three layers of the AI stack are a layered architecture, with the same rules about what may call what |
| Fundamentals + AI Engineering | Evolve in small steps | Her architecture chapter adds one piece at a time: context, guardrails, router and gateway, cache, agent patterns |
| The Hard Parts + AI Engineering | Data ownership | Who owns a table becomes who owns a dataset, an index, or an eval set |
| The Hard Parts + AI Engineering | Caching | Replicated caching between services; exact and semantic caching in front of a model |
| The Hard Parts + AI Engineering | Shared service or gateway | The reuse patterns (shared library, shared service, sidecar) are the options for a model gateway |
| The Hard Parts + AI Engineering | Analytical data and the data flywheel | Data mesh on one side, user feedback feeding the next model on the other |
The stamp coupling row is my favourite example of an old idea with a new price. Sending a whole customer record to a service that needs one field has always been wasteful. Sending a whole document to a model that needs one paragraph is the same mistake, except now every extra token is billed and makes the answer worse. See the fallacies of LLM systems for the evidence on long context.
For the observability row, AI agent observability shows what a trace of one agent request looks like. For the gateway row, Conway's law for AI covers who should own it.
4. What Each Book Gives You Alone
| Book | Only here | Read it first if |
|---|---|---|
| Fundamentals | Identifying and limiting architecture characteristics; cohesion, coupling and connascence; the catalog of styles from layered to microservices; diagramming, risk, negotiation and leadership | You have never had to justify an architecture to other people |
| The Hard Parts | Granularity disintegrators and integrators; decomposing a monolith and its database; data ownership; distributed transactions; the eight saga patterns; data mesh | You are splitting a system up, or debugging one that was split badly |
| AI Engineering | How foundation models work and are sampled; evaluation methodology; prompt engineering and prompt attacks; RAG and agents; finetuning; dataset engineering; inference optimization | You are putting a model into a product for the first time |
Two notes on AI Engineering. It is not a book about one vendor's API, so it has aged better than most writing on the subject. And its last chapter, "AI Engineering Architecture and User Feedback", is the one closest to the other two books: it builds a reference architecture one step at a time, which is how an architect would want it explained.
Get the AI engineering newsletter
Practical, engineering-first breakdowns on AI agents, LLMs, and breaking into the field — plus the free roadmap PDF when you join. No spam.
By subscribing you agree to receive emails from AI Engineer Insights. Unsubscribe anytime. See our Privacy Policy.
5. What None of Them Teach
Six things an AI architect deals with are outside all three books. That is not a criticism: books take years, and some of this didn't exist when they were written. You learn these from documentation and from running systems.
| Topic | Why the books don't cover it | Start here |
|---|---|---|
| MCP and agent protocols | The books predate most of the tooling | What is MCP? |
| Durable, long-running agents | Sagas cover long transactions, not agents that run for hours and must resume | Agent runtimes explained |
| AI governance and the EU AI Act | None of the three is a compliance book | AI guardrails |
| GPU capacity and FinOps | Inference optimization is covered; capacity planning and budgets are not | LLM routing |
| On-call for probabilistic systems | Runbooks assume the same input gives the same output | The fallacies of LLM systems |
| AI product sense | Knowing which problems a model should not be used for | Conway's law for AI |
6. Reading Order: Two 12-Week Plans
Which book you start with depends on which half you already have. Both plans end the same way: build something and write the decisions down.
If you are a software architect moving into AI
You have read the two architecture books, or lived them. Start with the model.
| Weeks | Read | Why |
|---|---|---|
| 1–2 | AI Engineering ch. 1–2 | What a foundation model is and why its output varies |
| 3–4 | AI Engineering ch. 3–4 | Evaluation. Read these as a book about fitness functions |
| 5 | AI Engineering ch. 5 | Prompting, including prompt injection |
| 6–7 | AI Engineering ch. 6 | RAG and agents |
| 8 | AI Engineering ch. 10 | The reference architecture: this is the chapter closest to your day job |
| 9–10 | AI Engineering ch. 7–9, skimmed | Finetuning, datasets and inference: know what exists and what it costs |
| 11 | The Hard Parts ch. 11–13, reread | Workflows, sagas and contracts, this time with agents in mind |
| 12 | Build | One AI feature with an eval suite in CI and an ADR for each big choice |
If you are an AI engineer growing into architecture
You can ship a RAG feature. What you are missing is the language for defending a design and the patterns for when it has to be split across teams.
| Weeks | Read | Why |
|---|---|---|
| 1–2 | Fundamentals ch. 1–5 | Architectural thinking, modularity, and how to pick characteristics |
| 3 | Fundamentals ch. 6–8 | Fitness functions, scope, components. You already write evals; this is the same habit |
| 4–5 | Fundamentals, the styles chapters | Read event-driven and microservices closely; skim the rest |
| 6 | Fundamentals, the chapters on decisions and risk | ADRs and risk storming |
| 7–8 | The Hard Parts ch. 1–4 and 7 | Coupling, quanta and service granularity |
| 9 | The Hard Parts ch. 8–10 | Reuse patterns, data ownership, data access |
| 10 | The Hard Parts ch. 11–13 | Orchestration versus choreography, sagas, contracts |
| 11 | The Hard Parts ch. 14–15 | Analytical data and trade-off analysis |
| 12 | Apply | Write ADRs and fitness functions for an AI system you already run |
If you are earlier than either, start with the AI engineer roadmap and come back to this list at phase 3.
Go deeper on AI engineering books and the AI architect reading map with AI
Ask your next question with this article as context. Opens in a new tab; you can edit the question before sending.
7. Frequently Asked Questions
What is the best AI engineering book?
For building products on foundation models, AI Engineering by Chip Huyen (O'Reilly, December 2024) is the most complete single book. It covers evaluation, prompt engineering, RAG and agents, finetuning, dataset engineering, inference optimization and a reference architecture. Pair it with a software architecture book if you design systems and not just features.
Do AI engineers need to learn software architecture?
Yes, once a feature has to run in production alongside other systems. Trade-off analysis, coupling, data ownership, contracts and distributed workflows all apply to AI systems, and the architecture books explain them more thoroughly than AI books do. Fundamentals of Software Architecture is the place to start.
What is an AI architect?
An AI architect designs systems in which some components are AI models. The role combines classic software architecture (characteristics, coupling, data, trade-offs) with AI engineering (evaluation, retrieval, agents, inference cost). There is no standard definition yet, which is why a reading map across both fields is useful.
What is the difference between a fitness function and an eval?
In practice, very little. A fitness function is an automated check that an architecture still has a property such as low coupling or acceptable latency. An eval is an automated check that an AI feature still produces acceptable output. Both run in CI and both exist to stop a property from drifting unnoticed.
In what order should I read these three books?
Software architects should read AI Engineering first, since they already know the other two topics. AI engineers should read Fundamentals of Software Architecture, then The Hard Parts. Read The Hard Parts after Fundamentals in either case, because it builds on it.
8. Sources
- • Mark Richards, Neal Ford — Fundamentals of Software Architecture, 2nd Edition (O'Reilly, March 2025)
- • Neal Ford, Mark Richards, Pramod Sadalage, Zhamak Dehghani — Software Architecture: The Hard Parts (O'Reilly, October 2021)
- • Chip Huyen — AI Engineering (O'Reilly, December 2024)
The mapping between the books is our own reading, not something the authors published. Chapter names and publication dates checked against the publisher's listings on November 16, 2026. No affiliate links.