Enterprise Agent Architecture: The Case for a Fifth Architecture Domain for the Agentic Enterprise

Enterprise architecture frameworks — TOGAF, the Zachman Framework, FEAF, SABSA, COBIT, and the security guidance codified by NIST — share a foundational assumption: that the enterprise architects, owns, and governs the systems it builds, and that the actors operating those systems are either human employees or deterministic software under configuration control. The arrival of autonomous AI agents violates this assumption. An agent is neither a system the enterprise fully specifies nor a person it can hold accountable; it is a non-human actor that holds delegated authority, decides its behavior at runtime, and composes tools it was never explicitly granted in combination. This paper argues that existing frameworks cannot model such an actor by extension, because the gap is structural rather than incidental: the frameworks describe what the enterprise *builds and runs*, while the agent introduces a *workforce* the enterprise must govern. We propose **Enterprise Agent Architecture (EAA)**, a fifth architecture domain that sits alongside the classical domains and is decomposed into four layers — Agent/Workforce, Capability, Control Plane, and Governance. We position EAA against the prior frameworks, show why each cannot absorb the agent as a sub-component, and outline the implications for architecture review boards, the division of authority between security and architecture leadership, and enterprise adoption.Version 2 (2026-07-05): corrected the characterization of SABSA's risk model in Sections 2 and 5. SABSA models trust as a graded, per-domain property and can represent insider threat; the paper's thesis (that SABSA provides no runtime-behavior layer for an authorized actor subverted through its inputs) is unchanged. No other substantive changes.