ARPE-IoT: Adaptive Recovery and Runtime Policy Enforcement for Resilient Cloud–Edge IoT Workflows

Abstract

Agentic artificial intelligence can improve Internet of Things (IoT) workflow resilience by reasoning over runtime context and generating recovery plans across devices, edge gateways, cloud services, and enterprise applications. However, autonomously completed recovery actions may still violate authorization, privacy, data-residency, actuator-safety, cost, or operational constraints. This paper proposes ARPE-IoT, an Adaptive Recovery and Policy Enforcement Framework for Agentic IoT Workflows. ARPE-IoT treats agent-generated recovery actions as proposals that must pass hard runtime policies before execution. Permissible actions are ranked using soft operational utility, executed under a bounded recovery budget, validated against workflow and physical-state conditions, audited, and escalated to humans when autonomous resolution is unsafe. We implemented a lightweight Python simulator with temperature and occupancy sensors, two edge gateways, two cloud services, an HVAC actuator, a policy engine, and a constrained recovery planner. Across 10,800 simulated executions covering normal operation and eight injected failure classes, ARPE-IoT achieved a 100% safe handling rate, consisting of 82.2% autonomous resolution and 17.8% controlled safe-state escalation. It reduced policy violations from 22.2% for unconstrained agents and 33.3% for rule-based recovery to 0% in the evaluated scenarios, with 47.7 ms mean latency.