There are two primary pillars upon which this argument is based. The first is the existence of hierarchical computer systems. Although this cannot be taken as proof that computer systems must be composed of levels, it gives strong empirical evidence that achieving complex behavior is much easier when components in a level provide an abstraction for components in higher levels.
The second pillar is Herb Simon's stability argument. He claims that complicated systems must be built out of relatively few subparts (and the subparts out of few subsubparts) to have a chance of succeeding. This is based on the assumption that the probability of a part failing is proportional to the number of components in the part. Therefore, a hierarchy is implicitly defined with the primitive components at the lowest level, assemblies of components at the next level, and so on until the entire system is constructed.
The primary use of levels in intelligence is to create abstractions to some degree, so higher levels can utilize a particular level without concerning itself with the details of implimentation. The completeness of this abstraction defines the strength of the level. The behavior of a strong level can be completely determined by the state of the system at that level, whereas a weak level would allow lower-level phonenomena to influence behavior.
A characteristic of level-based design is that each successive level is larger and slower than the lower levels. Since components at a particular level are composed of several components at lower levels, the size of the components increases geometrically with level. Similarly, since a component depends on the computation and interaction of its lower-level components, the amount of time a component takes to compute a function increases geometrically with level as well.
The assertion that intelligent systems, and in fact all complex systems, must be composed of levels is based on empirical and theoretical results. The ramifications of this claim can be used to help explain human cognitive architecture and design artificial architectures.