Clustering belief functions is not easy because of uncertainty and the unknown number of clusters. To overcome this problem, we extend agglomerative algorithm for clustering belief functions. By this extended algorithm, belief distance is taken as dissimilarity measure between two belief functions, and the complete-link algorithm is selected to calculate the dissimilarity between two clusters. Before every merging of two clusters, consistency test is executed. Only when the two clusters are consistent, they can merge, otherwise, dissimilarity between them is set to the largest value, which prevents them from merging and assists to determine the number of final clusters. Typical illustration shows same promising results. Firstly, the extended algorithm itself can determine the number of clusters instead of needing to set it in advance. Secondly, the extended algorithm can deal with belief functions with hidden conflict. At last, the algorithm extended is robust.