drobiazkin/ai-agent-architecture

This repository is a free, openly published engineering book and companion course about how large language model systems and AI agents actually work under the hood, written by a single author, Aleksandr Drobiazkin, in both English and Russian editions. It is not a collection of prompt writing tips. Instead it aims to teach engineers the underlying mechanisms, costs, and tradeoffs involved in building and running these systems themselves, so they are not entirely dependent on renting access to a handful of hosted AI providers. The book is organized into eight parts and twenty six chapters, covering topics like how text gets broken into tokens, how attention and different model architectures work, prompt and context design, retrieval augmented generation and its many variations, how AI agents are built as a kind of distributed system, integrating with tools through the Model Context Protocol, running these systems in production including security and cost concerns, and finally how to fine-tune, deploy, and even pre-train models on one's own hardware. Every major idea in the book follows a consistent template describing the context, the problem, the tradeoffs, the solution, a worked example, and related patterns, a structure borrowed from established software design pattern writing. The material is aimed at engineers who already understand distributed systems concepts and want to treat language models and agents as another architectural component to reason about rather than as something mysterious. Readers can browse the book chapter by chapter as plain text files in the repository, or build a single combined web page or PDF version themselves using the included build tools. A separate, practical companion course walks through building a real system across nine hands-on modules, from running a model on your own hardware through retrieval, agents, evaluation, and fine-tuning. The book's written text is released under the CC BY-SA 4.0 license, meaning it can be shared and adapted as long as credit is given and any reworked version is shared under the same terms, while the accompanying code in the repository is released separately under the MIT license. The author funds ongoing revisions through voluntary cryptocurrency donations rather than a paywall or subscription.