MKLab
AI Agents: The Definitive Guide [Koenigstein] - Printable Version

+- MKLab (https://mklab.gr)
+-- Forum: [INDEX] (https://mklab.gr/forumdisplay.php?fid=1)
+--- Forum: ARTFICIAL INTELLIGENCE (AI) (https://mklab.gr/forumdisplay.php?fid=5)
+---- Forum: BOOKS (https://mklab.gr/forumdisplay.php?fid=32)
+----- Forum: NEW PUBLICATIONS (https://mklab.gr/forumdisplay.php?fid=110)
+----- Thread: AI Agents: The Definitive Guide [Koenigstein] (/showthread.php?tid=1752)



AI Agents: The Definitive Guide [Koenigstein] - mklabgr - 08-30-2026

AI Agents: The Definitive Guide
Book:AI Agents: The Definitive Guide
Author: Nicole Koenigstein
Publication date: September 2026
Publisher: O’Reilly Media, Inc.
Level: Intermediate to advanced
Length: 378 pages (O'Reilly Media)

Summary

AI Agents: The Definitive Guide is an engineering-focused treatment of how to move AI agents beyond impressive prototypes into reliable production systems. Koenigstein begins with the conceptual transition from static LLM applications to stateful agents, using finite and hierarchical state machines to explain how agents maintain state, reason about goals, and determine their next actions. The book then develops major agent architectures—including reasoning-and-action patterns, planning, reflection, hierarchical agents, supervisor architectures, and multi-agent “swarms.” It also explores more advanced ideas such as reinforcement learning, search, test-time compute, Monte Carlo Tree Search, and techniques for improving agents through experience. 

A major strength of the book is its emphasis on the engineering infrastructure surrounding the LLM rather than treating the model itself as the entire agent. Koenigstein discusses model selection, mixture-of-experts and reasoning models, LoRA, quantization, structured outputs, Pydantic, tool calling, and the Model Context Protocol (MCP). She then addresses the difficult problems that emerge when agents enter production: sandboxing tool execution, Docker-based isolation, monitoring, fallbacks, inference infrastructure, evaluation, benchmarking, memory management, and computational cost. Agent memory is treated as an architectural component, with distinctions among episodic, semantic, procedural, short-term, and long-term memory and discussion of how memory should be organized across multi-agent systems. 

The final sections place particular emphasis on reliability and security. Agents need evaluation not only before deployment but continuously once they operate in real environments. The book therefore covers behavioral stress testing, production traces, custom benchmarks, long-horizon reasoning evaluation, threat modeling, red teaming, tool governance, layered defenses, and secure execution boundaries. Overall, the central message is that production-grade agentic AI is primarily a systems-engineering problem: useful agents require careful orchestration of models, state, memory, tools, infrastructure, evaluation, security, monitoring, and cost management rather than simply better prompting. 

Key takeaways
  • An AI agent is a system, not merely an LLM. State, planning, memory, tools, execution environments, monitoring, and feedback mechanisms are fundamental parts of the architecture.
  • Production reliability matters more than impressive demos. The book focuses heavily on contracts, structured outputs, testing, observability, fallbacks, sandboxing, and evaluation.
  • MCP and tool integration are becoming important infrastructure. Agents increasingly need standardized ways to discover and interact safely with external tools and services.
  • Security and evaluation must be architectural concerns from the beginning. Autonomous tool use increases the consequences of mistakes, making isolation, governance, red teaming, and continuous evaluation essential. 

Overall: This looks particularly useful for developers who already understand LLMs and basic agent frameworks and want to understand how serious agent systems should actually be engineered and deployed. It is less a “build your first chatbot” book and more a guide to the architecture and operational discipline required for dependable autonomous AI. 

O’Reilly book page