AGENT BEHAVIOR ENGINEERING: Building AI Agents That Know When to Act, Adapt, Ask and Stop - Couverture souple

Thomas, Nicholas

Couverture souple

ISBN 13 : 9798173238177

Editeur : Independently published, 2026

Synopsis

The next frontier of AI agents isn’t making them more intelligent. It’s engineering what they do with that intelligence.

AI agents are rapidly moving from generating answers to taking actions—using tools, remembering information, making decisions, delegating work, and interacting with real systems. But as agents gain autonomy, a harder engineering problem emerges:

How do you make an agent know when to act, when to adapt, when to ask, and when to stop?

Agent Behavior Engineering introduces a practical engineering discipline for building AI agents whose behavior is intentional, bounded, observable, testable, recoverable, and adaptable.

The central idea is simple but fundamental:

The model is not the behavior.

An agent’s behavior emerges from the interaction of its context, memory, tools, authority, state, runtime, environment, and evaluation mechanisms. A capable model can therefore make a reasonable decision and still produce the wrong real-world outcome.

This book shows developers, architects, AI engineers, technical leaders, and engineering teams how to design the system around the model—not just the model itself.

Inside, you’ll learn how to:

• Engineer the context behind every agent decision
• Decide what an agent should remember—and what it should forget
• Design safe tool contracts, permissions, authority, and side effects
• Define Behavioral Contracts using MUST, MAY, MUST ASK, MUST WAIT, MUST NOT, and MUST STOP
• Engineer decision boundaries for Act, Ask, Wait, Refuse, Stop, and Adapt
• Build stateful, durable agent runtimes that survive retries, failures, pauses, and recovery
• Understand behavioral failures including goal drift, specification gaming, prompt injection, memory poisoning, and authority escalation
• Debug autonomous systems using trajectories, decision provenance, replay, behavioral diffs, and the Behavioral Stack Trace
• Engineer safe multi-agent delegation, handoffs, shared state, trust, and collective behavior
• Turn Behavioral Contracts into behavioral tests and regression suites
• Detect behavioral drift and evaluate agents continuously in production
• Use runtime evidence to progressively increase, restrict, or revoke autonomy

Rather than focusing on one framework or vendor, the book develops durable engineering principles that apply across modern agent architectures.

At its center is a new way of thinking about autonomous systems:

Models create possibilities. Tools create consequences.
Capability determines what an agent can do. Behavior engineering determines what it should be allowed to do.

Through production scenarios, engineering patterns, architectures, failure cases, practical checks, and five foundational behavioral models, the book takes the reader from building an agent to engineering a dependable autonomous system.

If you already know how to build an AI agent, this book addresses the questions that begin after the first serious agent reaches production.

Because production agents must do more than reason well.

They must use power responsibly.
They must behave correctly under uncertainty.
They must survive failure without repeating consequences.
They must expose enough evidence to be debugged.
And they must earn autonomy through demonstrated reliability.

Agent Behavior Engineering is about turning capable AI agents into dependable engineering systems.

Build intelligence. Engineer behavior. Earn trust.

Détails de l'édition

• Éditeur: Independently published
• Date d'édition: 2026
• Langue: anglais
• ISBN 13: 9798173238177
• Reliure: Broché
• Nombre de pages: 327
• Coordonnées du fabricant: Printed by Amazon on behalf of the author; https://www.amazon.fr/hz/contact-us; ; c/o Amazon Media EU S.à.r.l., 38 Avenue John F. Kennedy; Luxembourg; L-1855; Luxembourg