Exploring the configuration space of radio frequency (RF) demodulation pipelines to find robust operating regimes
is challenging because simulation runs are computationally expensive and the response surface can contain sharp
failure boundaries. A naïve uniform grid of experiments wastes resources by sampling regions where the response
is already well understood. In this paper we demonstrate a probabilistic, agentic sweep strategy using Gaussian
process (GP) surrogate models to adaptively select candidate points. The GP provides predictive mean and
uncertainty estimates that allow sampling to focus on regions where the model is both uncertain and near the
robustness threshold. Using a synthetic RF demodulation benchmark we show that GPguided sampling can
discover robust operating modes with far fewer evaluations than uniform sampling. Boundary heatmaps and
uncertainty maps illustrate how the surrogate model rapidly localises failure rims. Our results support the use of
probabilistic active learning for efficient robustness characterisation and RF mode recovery.