Production-Grade Agentic AI

Beschreibung

Artificial intelligence is undergoing a major shift—from standalone models to connected, autonomous systems that can reason, act, and collaborate. This book explores that transformation, showing how intelligence is no longer confined to individual models but emerges from the systems that link them together. It also looks beyond language models, pointing toward the next frontier: world models. Designed for practitioners moving from experimentation to production, the book offers a clear and practical view of how modern agentic AI systems should be built, evaluated, and governed.

The book begins with the foundations of machine learning and highlights the limitations of large language models, including challenges with memory, action, coordination, and hallucination. From there, the book introduces the key architectural layers needed to make AI truly useful, starting with retrieval-augmented generation. It then presents a full framework for agentic systems, covering agent behavior, memory design, tool usage, and multi-agent collaboration. Critical topics such as interoperability standards, reliability measurement, prompt-injection risks, and sustainability through semantic caching are explained in a practical, implementation-focused way. Real-world applications are illustrated through enterprise use cases in areas like supply chains, document processing, and voice agents, along with guidance on data design and regulatory compliance.

Looking ahead, the book explores continual learning, advanced memory systems, and devotes a full chapter to world models — the frontier many researchers see as the next paradigm after LLMs. Throughout, optional deep dives take readers further into the material, including interviews with practitioners building agentic systems and open standards today. Woven through the technical chapters are three near-future scenes set in 2036, which put the book's arguments to work in situations a practitioner can recognize. Combining technical depth with real-world insight, it equips readers to build AI systems that are not only powerful, but also reliable, secure, and future-ready.

What you will learn:

· Design and implement multi‑agent architectures using proven patterns such as orchestrator, peer‑to‑peer, and hierarchical models.

· Measure and reduce hallucinations using the THS framework, FCD/FGR metrics, and the HalluBench‑310‑v3 benchmark.

· Secure agentic systems against prompt injection using a five‑layer defense architecture and the guardian agent pattern.

·  Classify an agentic system under the AI Act and understand where liability actually lands between model provider and deployer.

·  Evaluate world models — JEPA, Genie, Cosmos, GAIA — and recognize where a world model adds grounding that an LLM planner alone cannot provide.

Artificial intelligence is undergoing a major shift—from standalone models to connected, autonomous systems that can reason, act, and collaborate. This book explores that transformation, showing how intelligence is no longer confined to individual models but emerges from the systems that link them together. It also looks beyond language models, pointing toward the next frontier: world models. Designed for practitioners moving from experimentation to production, the book offers a clear and practical view of how modern agentic AI systems should be built, evaluated, and governed.

The book begins with the foundations of machine learning and highlights the limitations of large language models, including challenges with memory, action, coordination, and hallucination. From there, the book introduces the key architectural layers needed to make AI truly useful, starting with retrieval-augmented generation. It then presents a full framework for agentic systems, covering agent behavior, memory design, tool usage, and multi-agent collaboration. Critical topics such as interoperability standards, reliability measurement, prompt-injection risks, and sustainability through semantic caching are explained in a practical, implementation-focused way. Real-world applications are illustrated through enterprise use cases in areas like supply chains, document processing, and voice agents, along with guidance on data design and regulatory compliance.

Looking ahead, the book explores continual learning, advanced memory systems, and devotes a full chapter to world models — the frontier many researchers see as the next paradigm after LLMs. Throughout, optional deep dives take readers further into the material, including interviews with practitioners building agentic systems and open standards today. Woven through the technical chapters are three near-future scenes set in 2036, which put the book's arguments to work in situations a practitioner can recognize. Combining technical depth with real-world insight, it equips readers to build AI systems that are not only powerful, but also reliable, secure, and future-ready.

What you will learn:

· Design and implement multi‑agent architectures using proven patterns such as orchestrator, peer‑to‑peer, and hierarchical models.

· Measure and reduce hallucinations using the THS framework, FCD/FGR metrics, and the HalluBench‑310‑v3 benchmark.

· Secure agentic systems against prompt injection using a five‑layer defense architecture and the guardian agent pattern.

·Classify an agentic system under the AI Act and understand where liability actually lands between model provider and deployer.

·Evaluate world models — JEPA, Genie, Cosmos, GAIA — and recognize where a world model adds grounding that an LLM planner alone cannot provide.

 Who this book is for:

AI engineers, Data science practitioners, and software developers building or evaluating agentic systems; technical architects designing multi-agent pipelines; and the technology and compliance leaders — CTOs, heads of AI, risk and governance officers — who must decide where to deploy such systems and answer for them.

Autor*in

Diego Gosmar

Themen in »Production-Grade Agentic AI«

Agentic AI Multi-Agent Systems Large Language Models Retrieval Augmented Generation Prompt Injection Model Context Protocol

Stimmen zu »Production-Grade Agentic AI«

Details

Link teilen