Current logical approaches to describe intelligent agents tend to be over-abstract rather than being practical. In contrast, hybrid architectures in the software engineering tradition, such as INTERRAP [9], tend to neglect formal methods, leaving a gap between theory and practice. We present the Coop calculus, a language for concurrent, continuous inference processes, as a means to bridge this gap. Coop presents a declarative account of layered reasoning in a rational and logical setting, but abandons a rigid, cycle-oriented view to obtain an operationally reactive agent core.