We present a two-level coalition formation approach based on a centralized optimization model on the upper level, and a distributed agent-negotiation model on the lower level. This approach allows us to balance agent self-interests against a high joint utility. Experimental results show that the two-level coalition formation mechanism will increase not only the overall utility of the coalition, but also the individual utility of most participating agents. The results also suggest it is better for the agents to be partially cooperative rather than either fully cooperative or self-interested in our setting.