A short, opinionated reading list for AI solution architects. Everything here is either free or a book. Each entry says what it is good for and when to read it. Links were opened on 13 October 2026.
Two rules for using it. Prefer primary sources — the people who built or measured the thing — over commentary about them. And read with a project in hand: a guide read while designing something real is worth five read in the abstract.
If you only read ten things#
| # | Resource | Why this one |
|---|---|---|
| 1 | Building effective agents — Anthropic | The clearest statement of “start simple”: workflows versus agents and the five workflow shapes. Twenty minutes; rereads well |
| 2 | What We Learned from a Year of Building with LLMs — six practitioners | Tactical, operational and strategic lessons from people who shipped. The best single field report |
| 3 | Patterns for Building LLM-based Systems and Products — Eugene Yan | Evals, RAG, fine-tuning, caching, guardrails, feedback — each as a pattern with trade-offs |
| 4 | Your AI Product Needs Evals — Hamel Husain | Why evaluation is the centre of the work, with a worked example |
| 5 | AWS Well-Architected Generative AI Lens | A complete review checklist by pillar. Useful on any cloud |
| 6 | Azure Well-Architected Framework: AI workloads | Design principles with the trade-offs written out |
| 7 | Microsoft Cloud Adoption Framework: AI strategy | Use-case identification and the build-or-buy ladder, for the whole organisation |
| 8 | FinOps for AI overview — FinOps Foundation | Cost drivers, KPIs and pricing models in finance’s language |
| 9 | NIST AI Risk Management Framework | The structure most governance programmes are built on |
| 10 | The C4 model — Simon Brown | How to draw architecture so that different readers each get a picture at their level |
Starting from zero#
| Resource | What you get |
|---|---|
| Neural networks — 3Blue1Brown | Visual intuition for how neural networks and transformers work, no mathematics assumed |
| Generative AI for Beginners — Microsoft | A free, lesson-by-lesson course with code |
| Anthropic Academy | Free courses on the API, prompting, MCP and agent building |
| DeepLearning.AI short courses | One-to-two-hour courses on single topics, often taught by the tool’s makers |
| Hugging Face Agents Course | A free, hands-on introduction to agents |
| Google Skills, AWS Skill Builder, Microsoft Learn | Each cloud’s own free training, mapped to its certifications |
Patterns and architecture#
| Resource | What you get |
|---|---|
| AWS Prescriptive Guidance: Agentic AI | A six-guide series: foundations, patterns, frameworks, serverless, multi-tenant, operationalising |
| Azure Architecture Center: AI and machine learning | Reference architectures with diagrams you can adapt |
| Google Cloud Well-Architected Framework: AI and ML perspective | Recommendations by pillar |
| Effective context engineering for AI agents — Anthropic | Why what goes into the context window is the main design lever |
| A practical guide to building agents — OpenAI | When to build an agent, orchestration patterns, guardrails |
| Model Context Protocol and A2A | The specifications themselves. Read the introductions; skim the rest |
| Architecture decision records | Templates and examples |
| The Architect Elevator — Gregor Hohpe | How architects connect the engine room to the boardroom |
Evaluation, quality and operations#
| Resource | What you get |
|---|---|
| Your AI Product Needs Evals — Hamel Husain | The mindset and a practical loop |
| Inspect — UK AI Security Institute | An open evaluation framework, with documentation that teaches the concepts |
| OpenTelemetry semantic conventions for generative AI | What to record about each model and tool call |
Security, risk and regulation#
| Resource | What you get |
|---|---|
| OWASP GenAI Security Project | The Top 10 lists for LLM applications and for agentic applications |
| MITRE ATLAS | A catalogue of real attack techniques against AI systems |
| Simon Willison on prompt injection | The longest-running clear account of why the problem is hard |
| EU AI Act — official text and the AI Act Explorer | The law, and a navigable version of it |
| NIST AI 600-1: Generative AI Profile | Risks specific to generative AI, mapped to actions |
| ISO/IEC 42001 | The certifiable AI management-system standard |
Cost and models#
| Resource | What you get |
|---|---|
| Provider pricing pages: Claude, OpenAI, Gemini | The only reliable source for a price. Read them before any business case |
| Artificial Analysis | Independent comparisons of model quality, speed and price |
| FinOps for AI overview | KPIs and practices for AI spend |
The state of the field#
| Resource | What you get |
|---|---|
| Stanford AI Index | The annual, evidence-based survey of the whole field |
| McKinsey: The State of AI | Yearly survey of enterprise adoption and impact |
| Thoughtworks Technology Radar | Twice-yearly practitioner view of what to adopt, trial, assess and avoid |
Books#
| Book | Read it for |
|---|---|
| AI Engineering — Chip Huyen (O’Reilly, 2025) | The most complete single book on building applications with foundation models |
| Designing Machine Learning Systems — Chip Huyen (O’Reilly, 2022) | The classical ML system lifecycle, still relevant |
| Designing Data-Intensive Applications — Martin Kleppmann | The foundations of every data system an AI design sits on |
| Fundamentals of Software Architecture — Mark Richards and Neal Ford | Architecture as trade-offs; the vocabulary of the job |
| The Software Architect Elevator — Gregor Hohpe | Working between engineers and executives |
| The Mom Test — Rob Fitzpatrick | How to ask questions that get true answers. The best short book on discovery |
| How to Measure Anything — Douglas Hubbard | Putting numbers on things people say cannot be measured |
| The Pyramid Principle — Barbara Minto | Writing so that the first page carries the argument |
How to keep up without drowning#
- Subscribe to release notes, not news: the changelog of the one cloud AI platform and the one or two model providers you actually use.
- Read deprecation pages monthly. They are the only AI news with a deadline.
- Re-run your own evaluation set on each major release. Fifteen minutes of your own data beats an hour of other people’s opinions.
- Pick three people who build things and read what they write. Ignore the rest.
- Distrust anything without a date. In this field, undated advice is expired advice.