Recent studies show that the performance of deep convolutional neural network (CNN) applied to Steganography is better than that of traditional methods. We propose NAS-Stego to solve the problem of static network structure of the model in steganography algorithm based on deep learning. Different from the existing steganography algorithm based on deep learning, NAS-Stego uses a controller which is a LSTM[18] to generate the architecture of the encoder. We use reinforcement learning and Monte Carlo to train that controller. We have conducted experiments on BOSSBase dataset, and the results show that NAS-Stego has achieved better performance. We use steganalysis analyzer StegExpose to test the anti-steganalysis capability of NAS-Stego, the experiments show that NAS-Stego has achieved good performance.