Overview
An enterprise AI agent differs from a single-turn chatbot because it iterates through multiple steps to pursue a goal, using memory, tools, reasoning, and verification along the way.1 Since those steps may interact with real systems and data, governance, observability, and trust controls need to be built into the workflow rather than added after deployment.1
AWS Reference Architecture
The AWS reference architecture described in the source has five layers: an API gateway for identity and policy enforcement; a five-phase agent runtime; a knowledge and state tier for memory and context; platform services for models and tools; and a cross-cutting trust and control framework for governance and safety.1 Its runtime phases are Understand, Plan, Retrieve, Reason, and Act and Observe.1
The architecture is aligned with the AWS Well-Architected Framework Generative AI Lens and emphasizes zero trust, defense in depth, separation of concerns, observability by default, and fail-safe defaults.1 The source identifies orchestration, memory, cost, and control as challenges that accompany enterprise-agent adoption, and argues for designs that are governed, observable, trustworthy, iterative, and cost-aware.1
Design Implications
The architecture frames agents as controlled systems rather than simply more capable models: an agent's loop and access to tools require explicit identity and policy enforcement, persistent context management, and ongoing observation and control.1