Reliable AI Agents : Guardrails, Evals, and the Deterministic Core - Engineering Trust. Red-teaming, invariants, injection defense, observability, and the assurance case, built around a real credit age

Vu Tri Cong Mobiluck - Code247. Ai

ISBN 13: 9798171349462

Editore: Independently Published Sep 2026, 2026

Riguardo questo articolo

Neuware.

Codice articolo 9798171349462

Dati bibliografici

• Titolo: Reliable AI Agents : Guardrails, Evals, and the Deterministic Core - Engineering Trust. Red-teaming, invariants, injection defense, observability, and the assurance case, built around a real credit age
• Editore: Independently Published Sep 2026
• Data di pubblicazione: 2026
• Lingua: Inglese
• ISBN 13: 9798171349462
• Rilegatura: Taschenbuch
• Condizione: Neu
• Dimensioni: 229 L x 152 A x 8 P (millimetri) 229 millimetri di larghezza per 152 millimetri di altezza per 8 millimetri di profondità
• Peso: 214 grammi

Descrizione articolo

Riassunto

Your agent works in the demo. Now make it safe enough to put money — and a regulator — behind.

When a wrong answer is an awkward email, "usually works" is fine. When it is money that does not come back, a limit breached, a regulator's question, "usually works" is a liability with good demos. This book engineers the difference. Its running system is Harbor Credit — an agent that extends trade credit to a small business's wholesale buyers — and across ten chapters it wraps that agent in the full assurance pipeline that turns an impressive prototype into a system you can certify.

Through ten chapters, each moving from a reproduced failure to tested code, you will build:

• A red-team harness and a six-type failure taxonomy — because agent failures are systematic, not bad luck


• Evals that report quality and safety as separate numbers, with a calibrated LLM-judge and a golden set as executable spec


• Input and output guardrails — deterministic, fail-safe rails at the boundary, the model free in the middle


• The deterministic core: business invariants in plain Python that have the last word over whatever the model proposes


• Business invariants elicited from catastrophe, sorted from preferences, and pinned to acceptance cases


• A threat model and the strongest security an agent can have — safe by construction, where even an obeyed injection cannot breach a limit


• Observability with byte-for-byte trajectory replay — you debug trajectories, not lines


• Graceful degradation — timeouts, circuit breakers, and a safe mode that never auto-approves


• Property-based and adversarial testing that certifies distributions, not single cases


• The assurance case — the auditable, evidenced argument, with disclosed residual risk, that the agent is safe to deploy

The governing idea, held for ten chapters: put the LLM at the boundary and the invariants in the core, so safety holds regardless of model quality — perfect, degraded, or jailbroken. Each chapter adds a research lineage (FMEA and fault trees to reference monitors, QuickCheck to the safety case), a failure catalog, and exercises. The companion repo has one dependency, thirty-three tests, and runs entirely offline.

Who it's for: senior engineers, and anyone deploying agents in high-stakes or regulated domains, who have built and dissected an agent (Books 2–3 or equivalent).

The AI Agent Series takes readers from first contact to professional mastery across seven volumes. This is Book 5: trust, engineered.