In recent years, learning-based low-light image enhancement methods have shown excellent performance, but the heuristic design adopted by most methods requires high engineering skills for developers, causing expensive inference costs that are unfriendly to the hardware platform. To handle this issue, we propose to automatically discover an efficient architecture, called progressive attentive Retinex network (PAR-Net). We define a new attentive Retinex framework by introducing the attention mechanism to strengthen structural representation. A multi-level search space containing micro-level on the operation and macro-level on the cell is established to realize meticulous construction. To endow the searched architecture with the hardware-aware property, we develop a latency-constrained progressive search strategy that successfully improves the model capability by explicitly expressing the intrinsic relationship between different models defined in the attentive Retinex framework. Extensive quantitative and qualitative experimental results fully justify the superiority of our proposed approach against other state-of-the-art methods. A series of analytical evaluations is performed to illustrate the validity of our proposed algorithm.