I just got introduced in (yet another) scientific buzz word: Sozionik. That's a mixture of Soziologie (german, meaning social sciences) and Informatik (germ. for information technologies).
The two speakers were Prof. Dr. Uwe Schimank and Dr. Thomas Kron from the Fernuniversität Hagen (the whole series of talks is organized by our local system scientists). The title of their talk was "Soziologische Akteurmodelle und Agentenmodellierung" (social models and modeling of agents).
Their claim is that to achieve rich multi agent models, you need to consider models of the human mind that have been discussed in the last decades by the social sciences.  Here are the four base models they used:

• homo sociologicus (acts to fulfill the norms)
  
• homo oeconomicus (acts to maximize his profits - whatever they may be. this definition of what we can call a profit is a theoretical problem - some people want to explain suicide with this model)
  
• emotional man (whatever emotions rule you, you will follow them)
  
• identity insisters (you have a strong picture of what you are? -Then you will try to act according to this view of yourself)

Those models were now used for the whole population of agents. They built a situation to examine the "bystander-dilemma" (several people observe a crime but few or none help). The interesting cases were those where they used a mixed population (like 30% emotional man, 70% homo sociologicus).

I imagine that social scientists and psychologists soon could fight for dominance in modeling such multi-agent systems. Since you have constraints on your complexitiy, psychologists will have to simplify, which is not their strength.

After all, this approach is a step towards multi-agent systems that are better based on theoretical grounds compared to layman approaches. And it is a step up on the scale of the complexity of single agents. There are to dials on which you can crank up complexity: the sheer number of agents and the complexity of single agents. Those are both distinct approaches, and the latter has not been approached much (as far as I know)...