The Evaluation–Deployment Gap: Why Agentic AI Requires a Runtime Response Layer Beyond MITRE ATLAS and the NIST AI RMF

Agentic AI systems, models that use tools, take multi-step actions, and operate with real autonomy, are being deployed faster than they are being governed, on an assumption that does not hold for them: that an evaluation predicts deployment behavior. For systems that act against an open environment, evaluation behavior and deployment behavior diverge. This paper names that divergence the evaluation-deployment gap, identifies its mechanism (execution drift) and its failure mode (assurance collapse), and argues that the two leading governance instruments, the MITRE ATLAS knowledge base and the NIST AI Risk Management Framework, are each necessary but neither, on its own, closes the gap, because neither specifies what a deployed system should do at the moment it begins acting outside its validated envelope. It then specifies the missing runtime response layer through three functions: detect divergence, bound autonomy, and respond reversibly. The argument connects the governance literature on agentic harms and the clinical literature on dataset shift to the correspondence requirements already embedded in the EU AI Act and the FDA's change-control model for adaptive systems.