Agents introduce causality, reflection, necessity and various other sub-components never to be found in purely stochastic completion engines. This is an improvement, but it does require breaking down what each "agent" needs to do. What are the "core components" of cognition?
That's why I claim that any sufficiently complicated cognitive architecture contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of Immanuel Kant's work.