Researchers propose a dual-loop Agentic RF Intelligence architecture and show on a single Jetson Thor that millisecond wireless sensing and second-scale LLM reasoning can be decoupled and run concurrently

It presents a vision for Agentic RF Intelligence centered on a dual-loop architecture that separates fast, locally autonomous wireless sensing from slower agentic reasoning and orchestration. Rather than coupling sensing and reasoning at a single timescale, the architecture explicitly layers by timescale: the fast loop detects spectrum dynamics, interference, and signal sources, while the slow loop directs sensing and adapts network policies and resources. This is an architectural vision and design argument, supported by an on-device prototype and experimental observations rather than a large-scale comparative evaluation.

It provides an on-device prototype in which the full stack runs on a single NVIDIA Jetson Thor connected to a B200 Mini USRP. It realizes sensing, reasoning, and action on the same network-edge device, responding to deployment needs under strict latency, compute, and energy constraints. Prototype-level evidence showing the stack can run on a single edge device; the text does not report quantitative metrics such as power or throughput.

Experiments confirm the separation in timescales: WPFMs operate at millisecond latency, while LLM interactions take seconds to minutes. It provides direct latency observations supporting the dual-loop decoupling, showing the two workloads differ by orders of magnitude in response speed. Latency observations from the prototype, described as initial evidence; the text reports no sample sizes or statistical tests.

The fast loop remains operational during agentic reasoning, providing initial evidence for the feasibility of the decoupled architecture. It indicates that slow reasoning does not block fast perception, a key operating property for whether the dual-loop design holds. The text frames this as initial evidence, i.e., feasibility-level preliminary validation.

The work targets future 6G systems and Physical AI settings, applicable to deployments that need fast wireless sensing and slower reasoning orchestration at the network edge, for example detecting spectrum dynamics, interference, and signal sources and adapting network policies and resources accordingly. Its value lies in a reproducible architecture and on-device prototype path: the full stack runs on a single NVIDIA Jetson Thor connected to a B200 Mini USRP, the fast loop processes IQ samples with signal-processing tools and WPFMs, and a local LLM asynchronously interprets RF events, invokes tools, and steers sensing. Extensions outlined include memory-driven self-improvement, world models, network control, and multi-agent operation, making it relevant to researchers and engineers working on edge AI and wireless system architecture.

The current evidence comes from the abstract-level description, so the specific experimental setup, sample sizes, and statistical treatment remain unclear, and quantitative metrics such as power, throughput, and accuracy are not visible. The conclusion that the fast loop remains operational during reasoning is framed as initial evidence, and its robustness across different loads and hardware conditions remains to be seen. The outlined extensions toward memory-driven self-improvement, world models, network control, and multi-agent operation are unverified and remain open questions. Because this assessment is based only on the abstract, figures and body details are not included, and the above judgments may shift with the full text.