AGENTIC AI - Designing Autonomous AI Agents and Multi-Agent Systems (Paperback)

Paperback. For a decade the dominant question in applied artificial intelligence was how accurately a model could map an input to an output. A photograph went in and a label came out; a sentence went in and a translation came out. The model was a function, and the engineer's craft lay in fitting that function well. That era has not ended, but it has been subsumed. The question that now organises research laboratories, product roadmaps and graduate curricula alike is different in kind: not how well a model predicts, but how well a system acts - how it forms a goal, decomposes it, chooses instruments, recovers from its own mistakes, cooperates with other systems, and remains accountable to the people it serves.This book is about that shift. Agentic AI names a family of architectures in which a language model, or a society of them, is embedded in a loop with memory, tools, planning machinery and supervisory controls, so that the composite system can pursue objectives over extended horizons with limited human intervention. The engineering problems this creates are genuinely new. A model that hallucinates a citation produces a bad sentence; an agent that hallucinates a database table produces a bad migration. A model evaluated on a static benchmark is either right or wrong; an agent must be evaluated on trajectories, costs, side effects and recovery behaviour. A single model has no politics; a team of agents has delegation, disagreement, deadlock and emergent strategy.Who this book is forThe book was written for three overlapping audiences. The first is the senior undergraduate or postgraduate student taking a course in intelligent agents, applied machine learning or advanced software architecture, who needs a treatment that is rigorous about concepts and concrete about implementation. The second is the practising engineer or architect who must ship an agentic feature into production this quarter and would prefer to inherit a decade of multi-agent systems research rather than rediscover it through outages. The third is the researcher or technical leader who needs a map of the field - its results, its open problems and its genuine risks - before committing a programme to it.Readers are assumed to be comfortable with Python, with elementary probability and linear algebra, and with the basic idea of a transformer-based language model. No prior exposure to classical agent theory, reinforcement learning or distributed systems is required; where those fields are needed, they are developed from their foundations in the text.How the book is organisedPart I establishes what agency means and dissects the anatomy of an agentic system, including the cognitive architectures that recur in nearly every design. Part II treats the four core capabilities of an individual agent - planning, tool use, memory and self-improvement - and then the reinforcement-learning methods that convert experience into policy. Part III turns to multi-agent systems: communication, orchestration topologies, negotiation and market mechanisms, and the emergent behaviour of agent societies. Part IV is the engineering half of the book, covering frameworks, evaluation and observability, safety and security, deployment economics, and the hardware interfaces of embodied and edge agents. Part V closes with three complete capstone projects and a survey of the frontier.Every chapter follows the same rhythm. Concepts are introduced with a worked motivating example, formalised where formalism earns its keep, illustrated with figures and comparative tables, then reduced to runnable code. Each chapter ends with a short set of key takeaways and a graded set of exercises, some analytical and some requiring construction. Chapters are largely self-contained after Part I, so a reader with specific needs may enter the book at Shipping may be from multiple locations in the US or from the UK, depending on stock availability.…