As the Internet continues to expand into new application domains, there is growing demand for differentiated services that provide varying degrees of Quality of Service (QoS). Until recently, the approach for service differentiation in the Internet has focused on providing QoS guarantees to individual traffic flows. This per-flow model has not been widely embraced, largely due to the vast amount of state information that needs to be maintained in the network. As a consequence, the network community is currently redefining the notion of QoS for the Internet, leading to a new service model where guarantees are made to aggregate flows, rather than to individual flows. The notions of aggregate QoS which are being defined, e.g., by the Differentiated Services (DiffServ) working group in the IETF, are very different from traditional networking concepts and require novel algorithms for traffic control and provisioning. In our work in this area we seek to derive fundamental insights into the nature and delivery of QoS to traffic aggregates, i.e. aggregate QoS. A central feature of service models for aggregate QoS is that there is no requirement for users to specify the composition or the destination of traffic. This significantly complicates the provisioning of network resources and suggests that new, feedback-intensive traffic control algorithms must be developed to mitigate uncertainty about network loading. At present there is little or no theory available to guide the development of such algorithms. In this paper we lay out a research agenda for illuminating the fundamental issues associated with QoS for aggregate traffic. By providing a theoretical reference frame for aggregate QoS, one which provides tools and techniques for congestion control, capacity provisioning, and admission control, we complement ongoing standardization efforts for differentiated services in the Internet.