[ I shall concentrate on my science of interest (and what I know about) : psychology (cognitive-development psychology, in particular). Still, the principles I try to indicate should be of a general nature, true in-general for all sciences. This should also be applicable to artificial intelligence (computer science) as well. (Readers may be relieved to know only ONE sentence relates particularly to my theoretical view; still I do provide the reference to my main paper, at the bottom). ]

WE need the relatively simple, elemental, foundational things (good ones looked for and found) to provide the actual real foundation of good science and allow for real common understanding and giving us clear directions (for more and different, but related, research) and allowing for clear progress.  This, in fact, by-necessity or by-design, is the way things are and the way things work in other sciences -- look closely and see. 

IF you start out with ALL behaviors-of-interest (typically a LOT of behaviors for a psychological scientist, or an AI person) and these and their change (or causative) factors, both "BEING COMPLEX", THEN views of any behavioral-use/change-event-instance always then is said to involve several things, AND that with a lot of inherent uncertainties (here, there, and even "who knows where").  THEN you are relegated (doomed) to devising your own complex models "to explain things" (NOT A GOOD THING) (tell me: how could it be different?).  Such models so devised, no matter how seemingly inclusive and clear and logical and seemingly rational (and clever) and no matter how good some of the "explanations" seem to be:  I contend that this is never correct (or, even if somewhat useful for 'applied work', at least it is not correct in the long run); it is eventually clearly seen as deficient and the model beyond repair; if one continues to 'honor' the "system", one is "stuck". (There is a lot of writing by those with an interest in true artificial intelligence that say exactly this same thing, e.g. "Building Machines That Learn and Think Like People", by Lake et al , 2016)

In psychology, for example, in such a way: You will never be properly representing the CORE Subject of psychology (fundamental, foundational behavior-patterns-AND-environmental-aspects) -- no chance of this; you will actually not be clearly or usefully representing THE PERSON. This I maintain is necessarily true when EVERYTHING of interest to you is very multi-faceted (with multiple causations and "complex"). 

The the need is  for something relatively simple as a clear (elemental) part of the foundation of a science.

FIRST: Here's one 'good' example of a model in psychology:  In trying to explain everything (really!) about 'higher' cognition and cognitive processes, information-processing theories seemed very productive in the mid-1980s (e.g. John Anderson's ACT models). This finally became seen as an unacceptable model and as NOT the way things really are and AND thinking this way (basically BY-ANALOGY) did not provide good-enough (or even good) explanations and did not provide for good continuous progress (but some may still like i-p 'theory' (models) and disagree with this assessment; on the other hand, AI people are clearly "on my side" with regard to these models -- again, see the citation above).

For psychology, I submit we must find and recognize through observation (and reliably): BEHAVIOR PATTERNS -- the first recognition of the REALITY of this reality is seeing real patterns, FOUND directly, though direct concrete observations at appropriate times.   Soon through continuing work we can find the essentially "containing" SYSTEM (and not of our invention)(eventually enveloping all the most KEY things, aspects of behavior) BY continuing viewing actual behavior pattern changes in response to new aspects (or new patterns of aspects) of the environment (