In this paper, a distributed approach aimed at improving the quality of service in dynamic grid federations is presented. Virtual organizations (VO) are grouped into large-scale federations in which the original goals and scheduling mechanisms are left unchanged, while grid nodes can be quickly instructed to join or leave any VO at any time. Moreover, an agent-oriented framework is designed to observe and characterize past behaviors of nodes in terms of resource sharing and consumption, as well as to determine the trust relationships occurring between each pair of nodes. By combining trust and historical behaviors into a unified convenience measure, software agents are able to evaluate the (i) advantages of node's membership with VOs and (ii) whether a specific set of nodes is able to meet the actual requirements, in terms of resource sharing and consumption of a specific VO. The convenience measure has been exploited to design a fully decentralized, greedy procedure, aimed at controlling the grid formation process. Extensive simulations have shown that the coordinated and decentralized process of grid formation provides a powerful means to improve the overall quality of service of the grid federation.