Graph neural networks (GNNs) have demonstrated powerful performance in modeling high-order connectivity for collaborative filtering (CF) tasks. However, several challenges still remain unexplored in existing solutions: i) Diverse latent intents may influence users to choose certain items, which is widely ignored in representation learning; ii) The unreasonable combination of intent factors may ignore minor differences between users to some extent, which exacerbates the over-smoothing issue. To address above issues, we propose a novel Adaptive Disentangled Contrastive Collaborative Filtering (ADCCF) framework to distill fine-grained information from the entangled self-supervised signals. Specifically, we leverage cross-layer contrastive learning to guide the encoding process. Besides, ADCCF adaptively control the distance between each feature in the deep vector space, which can alleviate the over-smoothing problem. Extensive experiments on three real-world datasets indicate that our ADCCF outperforms various state-of-the-art recommendation methods.