Sources and Further Reading
Where the specific claims in this book come from, so you can check them. Grouped by what you might want to trace rather than listed alphabetically.
Where the specific claims in this book come from, so you can check them. Grouped by what you might want to trace rather than listed alphabetically.
Two general cautions. Published documentation about commercial AI systems is revised frequently, so a page may have changed since this book was written — check the date on whatever you find. And a figure reported by a company about its own system is evidence about that system, not about the technology in general.
The multi-agent research architecture in Chapter 11
Everything in Chapter 11 about how research work is divided among a lead agent and subagents — the orchestrator and worker arrangement, the separate citation pass, the 90.2 per cent improvement on internal research evaluations, the roughly fourfold and fifteenfold token figures, the observation that token use explains around eighty per cent of performance variation, the effort-scaling rules, the forty per cent improvement from a tool-description testing agent, and the full catalogue of failure modes — is drawn from:
- “How we built our multi-agent research system”, Anthropic engineering blog, June 2025. anthropic.com/engineering/multi-agent-research-system
- “When to use multi-agent systems (and when not to)”, Anthropic. A companion piece on the decision itself, and a good next read after Chapter 12.
The statement that multi-agent arrangements struggle with coding tasks because of high interdependency between the pieces of work is Anthropic's own, in the first of those.
Coding agents and subagents in Chapters 9 and 11
- Claude Code documentation, Anthropic, in particular the pages on subagents. This is the source for when to send a specialist and when to stay in the main loop, and for the existence of default limits on how deep delegation may nest and how many subagents may run at once. Those defaults are configurable and have changed over time, which is why this book describes them qualitatively.
General-purpose assistants in Chapter 10
The description of what happens when an assistant searches, uses a tool or remembers something is deliberately generic, because the details differ by product, by version and by a user's own settings. For specifics, the primary sources are the providers' own product and developer documentation:
- OpenAI platform and product documentation, for ChatGPT's search, tool use, code execution and memory features.
- Anthropic product and developer documentation, for the equivalent Claude features.
- Google documentation, for Gemini.
The code in Listing 1 uses LangChain's current agent interface. Library interfaces change between versions; check docs.langchain.com before building on it.
The epigraphs
- Andrew Ng, “AI is the new electricity” — Stanford MSx Future Forum, 2017, and repeated in his talks and writing since.
- Jensen Huang, on every company and every job having AI assistance — fireside chat opening SIGGRAPH 2024, Denver, 29 July 2024; reported on NVIDIA's own blog.
- Satya Nadella, on humans and swarms of AI agents — CEO Connect, Bengaluru, 2025.
- Sundar Pichai, on aligning AI systems to human values needing social scientists, ethicists and philosophers alongside engineers — interview on 60 Minutes, CBS, 2023.
Two epigraphs, at Parts II and V, are the author's own.
Data protection and standards in Chapter 15
Table 15-1 names instruments rather than summarising them, and the summaries there are orientation only. The authoritative texts are:
- GDPR — Regulation (EU) 2016/679.
- HIPAA and CCPA — United States federal health privacy legislation, and the California Consumer Privacy Act as amended.
- PIPEDA — Canada's Personal Information Protection and Electronic Documents Act.
- LGPD — Brazil's Lei Geral de Proteção de Dados.
- ISO/IEC 27001 — a voluntary international standard for information security management, published by ISO and IEC. Not a data protection law, and adopted by choice.
- The EU AI Act — Regulation (EU) 2024/1689, which post-dates most of the instruments above and is the most significant AI-specific regulation currently in force.
None of this is legal advice. Take proper advice for your own jurisdiction and use.
Governance and evaluation in Chapter 16
- AI Agent Governance: An End-to-End Guide to Evaluating, Governing, and Releasing AI Agents with ProofAgent, Dr. Fouad Bousetouane. The source for the ProofAgent governance framework, whose four axes — evaluation, context, compliance and governance, labelled E, Q, C and G — are the four questions in Chapter 16. It sets out how each axis is scored, how they are aggregated with limited compensation, the hard blockers, adversarial scenario design, control design and the release gate. Chapter 16 is a summary of it; nothing in Chapter 16 requires it.
- The ProofAgent Harness, the evaluation harness described there, is open source.
Foundational work referred to in passing
- The transformer — Vaswani et al., “Attention Is All You Need”, 2017. The paper that introduced the architecture behind every model in this book.
- Retrieval-augmented generation — Lewis et al., 2020, which named the pattern Chapter 4 describes.
- ELIZA — Joseph Weizenbaum, 1966. Worth reading about for how little machinery it took to be convincing.
- AlphaGo — DeepMind, 2016.
Where to go next
If you want to keep learning after this book, the most useful habit is not reading more books. It is reading the engineering blogs of the companies building these systems. They publish what worked, what failed and what it cost, in more detail than any secondary source will give you — and Chapter 11 exists because one of them did exactly that.
Beyond that: the documentation of whichever framework you choose, the release notes of whichever model you use, and twenty test cases of your own.