Building Complex Multi-Agent Systems Using Pattern Prompting A guide to building robust and secure GenAI applications using… (Tim OBrien) (z-library.sk, 1lib.sk, z-lib.sk)

【One-Line Pitch】 A pattern-first guide for experienced engineers who want to build secure, scalable GenAI and multi-agent systems using the architectural vocabulary they already know—GoF patterns, enterprise integration, and message brokers—rather than chasing framework churn. Best for architects and technical leads adding agentic AI to production infrastructure. 【Book Arc】 - **Opening (~0%–15%)**: Frames GenAI as a continuation of enterprise integration and big data, then builds a shared vocabulary by mapping GenAI terms (agent, tool, memory, multi-agent system) onto traditional IT equivalents (component, adapter, session state, system of record). - **Early (~15%–35%)**: Establishes the retrieval foundation—embeddings, semantic affinity, vector databases, chunking and overlap—and surveys the LLM landscape, positioning these as core knowledge for everything that follows. - **Middle (~35%–55%)**: Turns to tuning and architecture: the parameter dependency graph (chunk size, similarity function, temperature, max tokens), role-based temperature choices (judge, router, actor), and the case for RabbitMQ as the orchestration substrate for enterprise-grade GenAI. - **Late (~55%–80%)**: Moves into hands-on construction—building a first RAG app, wiring exchanges, queues, retries, fallbacks, and message history, plus data migration and ingestion pipelines (Airbyte, Pinecone) for preparing enterprise data. - **Ending (~80%–100%)**: Covers operational discipline—security (prompt injection, PII, data leakage), cost and token budgeting, prompt versioning and governance—and the Topologos workflow that generates, extends, and deploys multi-agent patterns as RabbitMQ manifests. 【Key Takeaways】 - **GenAI is enterprise integration wearing new clothes** (Early): The book's central thesis maps agentic patterns to microarchitectures of GoF and Fowler patterns, so existing orchestration and messaging skills transfer directly. - **A harmonized vocabulary is the real unlock** (Early): The terminology cheat sheet (agent→component, tool→adapter, long-term memory→system of record) lets you reason about architecture instead of memorizing tools. - **Retrieval quality is a tuning problem, not a model problem** (Early–Middle): Chunking strategy, overlap, similarity function, and max chunks returned form a dependency graph—change one and you must rebalance others. - **Temperature should follow component role** (Middle): Judges, routers, and actors favor low temperature for determinism; the book ties creativity settings to concrete architectural responsibilities. - **RabbitMQ provides the reliability LLMs lack** (Middle): Treating LLMs as unreliable API endpoints, the book uses exchanges, queues, dead-letter queues, retries, and fallbacks to build robust RAG and multi-agent flows. - **Prompts are code and need governance** (Ending): Versioning, review workflows, and treating prompts as first-class artifacts are presented as non-negotiable for production GenAI. - **Security and cost are design-time concerns** (Ending): Prompt injection, context-window data leakage, PII handling, and token budgeting are addressed as architectural decisions, not afterthoughts. - **Patterns can be generated, composed, and extended** (Ending): The Topologos workflow (/define, /deploy) lets you invoke canonical patterns like ReAct, compose them (ReAct + Tool Use), or define new ones such as Debate-and-Adjudicate. 【Reading Tips】 - **Deep-read the terminology mapping and pattern chapters** (Early–Middle): These are the book's differentiator; skimming them loses the reasoning framework that makes the rest coherent. - **Skim the introductory LLM/embedding survey if you already know RAG**: The vector database and embedding material is foundational but high-level; move quickly to the tuning and RabbitMQ chapters. - **Treat the parameter dependency graph as a working checklist**: When tuning your own system, revisit it to avoid fixing one parameter while silently degrading another. - **Use the worked examples as templates**: The ReAct, composed-pattern, and Debate-and-Adjudicate walkthroughs are the most directly reusable material for your own builds. - **Read the security and cost sections before you ship, not after**: Prompt injection and token budgeting are easy to defer and expensive to retrofit. 【Coverage Limits】 The excerpts cover the book's framing, retrieval fundamentals, tuning, RabbitMQ orchestration, and the Topologos workflow, but do not include full chapter text for the data migration, ingestion, or appendix pattern reference material. Specific code listings and complete worked examples are only partially represented.