A Taxonomy of Structural Defects in Autonomous LLM-Based Agent Systems

Autonomous LLM-based agent systems exhibit a class of failures that cannot be characterized as implementation bugs: they are structural defects arising from architectural decisions that violate the formal invariants required for deterministic, auditable, and epistemically valid operation. This paper presents the first formal taxonomy of such defects, derived from the HADD invariant set and grounded in anonymous instances from deployed systems. We identify and formally characterize seven defect classes: D1 Epistemic Staleness, D2 Non-deterministic Control Flow, D3 Post-hoc Filtering Without Causal Trace, D4 Unbounded Plan Synthesis, D5 Capability Precondition Bypass, D6 Single-Origin Belief Amplification, and D7 Supervisory Collapse Under Scale. For each class we provide a formal definition, the violated HADD invariant, an anonymous real-world instance, and a structural mitigation. We prove the classes are pairwise structurally distinct and complete with respect to the HADD invariant set. An audit checklist operationalizes the taxonomy for regulatory and engineering practice.