High-precision sensing systems often have multiple sensor modalities at their disposal, each trading measurement precision against acquisition time. Multi-modal sensing protocols are typically designed by hand and often sense more than necessary, leaving open the question of which sensor to query, where, and at what cost. We cast sensor selection as a modality-level action within the active inference framework: an agent maintains a generative model in which each sensor contributes a likelihood with its own precision and time cost, and selects actions by minimising expected free energy under a finite time budget. We demonstrate it on a low-order polynomial identification task. The inferred policy is shaped by how tightly time is budgeted: without a time penalty the agent keeps the precise sensor while it is most informative and releases it once the posterior has tightened. Under heavy time pressure, however, it abandons the precise sensor and relies on cheap, fast measurements. Its advantage is adaptive scheduling rather than uniform dominance. Under time pressure, non-myopic planning outperforms myopic selection by positioning the sensor at informative measurement locations.