Mechanically agitated granular matter often serves as a prototype for exploring the rich physics associated with hard-sphere systems, with an effective temperature introduced by vibrating or shaking. While depletion interactions drive clustering and assembly in colloids, no equivalent short-range attractions exist between macroscopic grains. Here we overcome this limitation and investigate granular cluster formation by using acoustic levitation and trapping. Scattered sound establishes short-range attractions between small particles, while detuning the acoustic trap generates active fluctuations. To illuminate the interplay between attractions and fluctuations, we investigate transitions among ground states of two-dimensional clusters composed of a few particles. Our main results, obtained using experiments and modelling, reveal that, in contrast to thermal colloids, in non-equilibrium granular ensembles the magnitude of active fluctuations controls not only the assembly rates but also their assembly pathways and ground-state statistics. These results open up new possibilities for non-invasively manipulating macroscopic particles, tuning their interactions and directing their assembly.