Agentic RF Intelligence: Multi-Timescale 6G Sensing and Reasoning with On-Device Foundation Models

Abstract

Future 6G systems share a central challenge with Physical AI: combining sensing, reasoning, and action on network-edge infrastructure in rapidly changing physical environments under strict latency, compute, and energy constraints. Addressing this challenge requires multi-timescale intelligence, combining fast perception that tracks wireless phenomena within milliseconds with slower reasoning that directs sensing and adapts network policies and resources over seconds. In this paper, we present a vision for Agentic RF Intelligence based on a dual-loop architecture that decouples fast, locally autonomous wireless sensing (e.g., detecting spectrum dynamics, interference, and signal sources) from slower agentic reasoning and orchestration. We provide an on-device prototype in which the full stack runs on a single NVIDIA Jetson Thor connected to a B200 Mini USRP. The fast loop processes IQ samples using signal-processing tools and wireless physical-layer foundation models (WPFM), while a local Large Language Model (LLM) asynchronously interprets RF events, invokes tools and steers sensing. Our experiments confirm the separation in timescales, as WPFMs operate at millisecond latency, while LLM interactions take seconds to minutes. The fast loop remains operational during agentic reasoning, providing initial evidence for the feasibility of this decoupled architecture. We conclude by outlining extensions toward memory-driven self-improvement, world models, network control, and multi-agent operation.