Traditional recommendation system has the problem of sparse user ratings and system scalability. This paper proposes a recommendation system based on intelligence multi-agent. At first, the cosine similarity measure has been used to handle user-item rating matrix, thus the initial neighbor set for target users can be gained. Then, user ratings have been mapped to relevant item attributes for generating user-attributes value preference matrix UPm of each user. Thus, user similarity can be computed based on UPm and rating sparsity has been alleviated simultaneously. The recommendation system of intelligence multi-agent makes calculating an online processing, and thus improves the system scalability. Experimental results show that the new system achieves a better accuracy in recommended convergence.