In this paper, we have outlined in broad brushstrokes the philosophy behind our multi-agent infrastructure for large-scale digital libraries, with an emphasis on the importance of having agents within that infrastructure that have knowledge and strategies for finding resources and forming agent teams. We have described task planning agents as a class of such agents, and have described our current efforts in developing the architecture, world and domain knowledge, and language for such agents - and in particular, for task planning agents that support query planning. While there are many unresolved issues in our design, we believe that we have a good starting point for developing such agents, and we have already begun to incorporate and evaluate prototype versions of our TPAs within the UMDL system.