A Layered Taxonomy of Agentic AI Failure Modes: Orthogonal Patterns and Three Corrective Primitives for Enterprise Governance

Abstract

As agentic AI systems transition from research prototypes to enterprise deployment, the absence of comprehensive failure mode frameworks impedes systematic risk assessment. This paper presents a practitioner taxonomy of failure patterns organized into a four-layer architectural hierarchy: Cognitive Core, Execution Interface, Interaction Boundary, and Adversarial Surface. We establish orthogonality criteria to minimize conceptual overlap, enabling structured risk quantification. Beyond enu- meration, we propose three corrective primitives—Causal State Fabric, Verified Bounded Execution, and Continuous Teleological Audit—as architectural interventions addressing the majority of catalogued failures. The Causal State Fabric, extending Semantic MVCC with provenance, causality, and uncertainty tracking, serves as the foundation upon which the other primitives depend. We argue that traditional quarterly gov- ernance cycles are mismatched to agentic failure dynamics, motivating continuous monitoring infrastructure. The taxonomy maps to regulatory frameworks including SR 11-7 model risk management and NAIC insur- ance AI guidance. This is a practitioner framework, not an empirically validated research contribution; we discuss limitations and validation requirements throughout.