Subsumption's behavior may be decomposed into three aspects: coherence, salience, and adequacy.
Subsumption Architecture agents may be considered coherent because they always are looking at their inputs and executing tbe best course-of-action based on them. In this sense they are Markov inspired: Markov advocated making devices whose sensory input was rich enough for them to uniquely decide the next best action, without resorting to an internal model of the world state.
Subsumption Architecture agents may be considered partially salient. Agents completely rely on their sensors to tell them what to do. If a sensor exists to interrupt the current task for one of a higher priority then Subsumption agents may pursue the new task. But if not, subsumption agent behavior may not be salient because they lack an internal model of the world from which to make conclusions. This deficiency has been recognized: one extension on the basic subsumption philosophy is the use of hormonal activity to provide some global state.
It is debatable whether or not Subsumption Architecture is adequate. True: it does provide realtime response to a complex and dynamic environment, but the response is limited to reflexive repertoire of actions. To increase its applicability, Subsumption has been augmented to create a new class of behavior-based architectures. An example behavior-based architecture is a planning and learning agent.
Go to a discussion of this property for multiple architectures.