Engineering self-adaptive systems is a particularly challenging problem. On the one hand, it is hard to develop the right control model that drives the adaptation; on the other hand, the implementation and integration of this control model into the target system is a difficult and an error-prone activity. Models@runtime is a promising approach to managing adaptations at runtime, as they provide higher levels of abstractions of both the running system and its environment. However, recent work mainly focuses on runtime mod-els that are causally connected to running systems and less attention is paid to how models can be used to develop and manage the control logic that drives runtime adaptations. In this paper we propose an alternative form of models@runtime as a reactive data-driven model centered around feedback control loops. Both the target system and the adaptation logic are represented as networks of message passing actors. Each of these actors represents a particular abstraction over the running system (sensors, effectors) and its control (analysis, decision). Moreover, the actors are also viewed as target systems themselves. This makes the feedback loops adaptable at runtime as well and permits us to build complex solutions with hierarchical layers of control loops. We discuss how this representation fits some of the requirements of models@runtime and helps to prototype a feedback control system on a concrete example extracted from ongoing validation case studies.