In this Letter, we adapt the direct search method to metasurface optimization. We show that the direct search algorithm, when coupled with deep learning techniques for free-form meta-atom generation, offers a computationally efficient optimization approach for metasurface optics. As an example, we apply the approach to optimization of achromatic metalenses. Taking advantage of the diverse dispersion responses of free-form meta-atoms, metalenses designed using this approach exhibit superior broadband performances compared to their multilevel diffractive counterparts. We further demonstrate an achromatic and wide-field-of-view metalens design.