PrivateAgent: Mobility-Aware AI Agents for Safe Private Networks

Abstract Private cellular networks increasingly support video analytics, connected workers, autonomous mobile robots (AMRs), and other industrial applications whose mobility and control deadlines vary widely. A single artificial-intelligence (AI) agent cannot safely manage all of these loops because perception, planning, networking, and governance operate at different time scales and execution locations. This paper presents PrivateAgent, a mobility-aware framework that classifies private-network applications as stationary, low-mobility, or high-mobility and maps their functions across device, site-edge, and cloud tiers. Hard real-time perception and actuation remain local; near-real-time planning and network optimization run at the edge; long-horizon learning runs in the cloud. Network-changing actions are represented as typed artifacts with explicit evidence, scope, authority, monitoring, and rollback requirements before execution. We detail representative use cases including perimeter video, machine vision, connected workers, forklifts, AMRs, and drones, and specify suitable agent families, latency targets, failure modes, and benchmark dimensions. We also instantiate the action-admission boundary with SafeFlow-Real50, a 50-scenario benchmark derived from four anonymized private-network traces. In a blind evaluation, the model selected the correct act/clarify/reject mode in 50/50 scenarios, an acceptable action in 45/50, and the preferred action in 43/50. The deterministic gateway detected all 400 injected corruptions versus 150 for schema-only validation. These results show that mobility should determine the real-time agent boundary, while independent validation should determine whether generated network actions may advance toward execution.