As an important coordination and cooperation method in multi-agent system, agent coalition mechanism has been receiving more and more attention. An efficient algorithm is needed for this topic since the number of the possible coalitions is exponential. This work proposes an improved ant colony optimization algorithm to find the optimal, task-oriented agent coalition in multi-agent system. Ants incline to choose those agents who cooperated well before to form coalitions, which realizes the acquaintance mechanism. The novel "inner hormone" can avoid the algorithm getting in the local minimum area easily. The results of contrastive experiment show that the algorithm in This work is robust, self-adaptive and very efficient.