In sequential recommendation (SR) systems, user behaviors are driven by various sophisticated intents during the decision-making process. In previous studies, sequences were aggregated in latent space to derive intent prototypes directly. However, these methods do not sufficiently explore higher-order transition patterns among items, leading to learned intents that capture only local correlations within sequences while lacking global collaborative information among sequences. To this end, we propose Global-Local Collaborative Learning for Intent-Aware Sequential Recommendation (CISR), which provides high-quality item embeddings for intent modeling and sequence modeling, thereby enabling the learning of more precise and informative intents. Specifically, we construct an adaptive item transition graph among all items, then seamlessly incorporate the global graph information into the item embeddings through mutual information maximization. In addition, to unify the semantic patterns of the sequence and intent, we propose a sequence encoder that includes local and global modules. The encoder maintains global collaboration while incorporating personalized features into the encoding process as relative position encoding. We enhance the integration of learned intents into the sequential model by maximizing the consistency of similar intents and the diversity of distinct intents. Extensive experiments on three public datasets demonstrate the superior performance and robustness of the proposed method.