Agentic LLM Compilation of Adaptive Sampling Policies for Low-Power IoT Sensing

Battery-powered sensors must sample sparingly to conserve energy while still reconstructing the signal faithfully between measurements, and the decision logic must fit microcontroller-class hardware. We use a large language model (LLM) as an offline policy compiler: given a natural-language task description and a compact rule grammar, it returns policies as a few numeric thresholds and integer periods, under 100 bytes. No weights are fitted and the model never sees the raw series, only eight summary statistics: the approach is training-free but statistics-informed. Policies come from an agentic loop: a generator LLM proposes candidates, a deterministic evaluator scores each exactly on the recorded trace, and an optimizer LLM refines them over eight rounds. We evaluate Claude Opus 4.8, Claude Fable 5, ChatGPT 5.4 and ChatGPT 5.5, five seeds each, on the Beijing PM_2.5 series against classical baselines, a budget-matched random search, and NSGA-II and ParEGO over the identical grammar. The loop improves on zero-shot generation and random search (pooled Wilcoxon p<10-5): within a 0.20 duty-cycle budget the best model's mean across five seeds reaches an RMSE of 21.79 against 25.09 for random search, and 15.98 against 20.57 within 0.30. The gain narrows at the tightest subsampling, which an interpolation-error bound explains through the signal's curvature.