Decoupling has become a new paradigm in Graph Neural Networks (GNNs) for its effectiveness and scalability. However, this paradigm still faces two several restrictions: unsatisfying propagation, caused by noisy or confused edges, could greatly degrade model performance; fixed aggregation schema with the same propagation steps and the same combination weights for each node limit achieving optimal performance. To address these problems, we propose a novel decoupled graph model named LA-DGNN based on label agreement message propagation and combine the intermediate feature after each propagation step as input. In our method, we decouple the graph model which trains a base predictor based on multi-layer perceptrons with a pre-step to propagate features and a post-step to propagate labels. We utilize an auxiliary label agreement model to generate proper edge weights to promote reliable propagation. When training the base predictor, we concatenate all intermediate features after each propagation step to make the model dynamically learn information of neighbors at different distances. Extensive experiments on five real-world datasets demonstrate that our method achieves superior performance over all baseline methods in terms of node classification accuracy.