Exhaustive parameter sweeps are the de facto method for benchmarking RF pipelines, but
they scale poorly with dimensionality[2]. For example, a 10-parameter grid with 10 points each
requires 1010 evaluations. Active learning promises to achieve comparable confidence using far
fewer samples by targeting the most informative points[1]. This paper constructs a synthetic
ground-truth generator for an RF performance field and compares random grid sampling to
a Gaussian process (GP) guided “agentic sweep.” We measure coverage of the true decision
boundary as a function of sample budget and explore how exploration versus exploitation balances affect performance. Results show that uncertainty-guided sampling achieves the same
classification coverage with approximately 3× fewer samples than random grids at 90% coverage thresholds, and that a modest amount of random exploration is beneficial. These insights
support using agentic sweeps for efficient characterization of RF demodulation pipelines.