Preface

A multi-agent system is never deployed to production simply because it succeeded in a controlled demo. Its genuine operational utility is tested when it ingests corrupted data, interacts with degraded third-party dependencies, and operates under strict regulatory obligations.

This volume dissects that real-world operating environment. In the domain of security, it maps the attack surfaces that emerge the moment a model is granted permission to query proprietary databases, invoke destructive tools, and delegate execution to sub-agents. Across healthcare, grid energy, financial markets, and logistics, it examines how these agentic capabilities interface with human oversight and audited business processes. The geopolitical analysis anchors the system within the global supply chain of frontier model labs, semiconductor fabs, cloud hyperscalers, and developer tooling—dependencies no systems architect can afford to ignore.

The three perspectives woven into each chapter—technical, executive, and academic—answer distinct questions anchored to the same empirical baseline. The technical lens details mechanics, invariants, and failure modes; the executive track evaluates cost-risk trade-offs and investment decisions; the academic track examines evidentiary strength and empirical methodology. They do not constitute three redundant retellings, but three coherent reading depths designed to inform decisive architectural judgment.

Developer tools and IDEs cited throughout the text represent snapshots of a volatile marketplace, not permanent endorsements. Architectural evaluation criteria transcend commercial brands: sandbox isolation, audit telemetry, spend caps, multi-model portability, and non-repudiation of agent actions.

Security is not an ornamental afterthought patched onto a finished system; it is the prerequisite that permits the system to continue running.

Sergio Perez Ruiz
Lima, 2026