We have taken Cederman’s assertion as a pretext for a journey through the world of artificial agents. From SAI to ComMod, we have drifted from positivism to constructivism, from a normal kind of science to a post-normal one. Despite the weaknesses and flaws pointed at in this chapter, the author, like Cederman, believes that Multi Agent Systems (MAS) offer a fantastic opportunity for the social, natural, and computer sciences to come together and engage in a ‘new kind of science’, much more challenging in epistemological terms than Wolfram’s original proposal (see Lissak and Richardson 2001). Unlike Kluver and colleagues (2003), we think that there is a way for social sciences to appropriate computer formalisms without falling into a reductionist and positivist stand. The enactive cognitive theory (Maturana and Varela 1980) and the interactionist theory (Minsky 1985) need to be re-visited in a transdisciplinary manner. They can provide a theoretical substance to a vast majority of MAS applications built with non-formal logic compliant agents. The validation of these models needs to be embedded into a constructivist perspective where designers, users, and stakeholders not only evaluate the simulated outcomes, but also participate in the modelling process itself (Granath 1991; Collectif ComMod 2005).

Progressing on to the constructivist path doesn’t mean that we have to discard the symbolico-cognitivist paradigm altogether. The strong, but limited, edifice proposed by SAI has shown undisputable capacities in producing replicable methods to test assumptions or benchmark findings on rational decision and action. For example, recent work from Castelfranchi (2001) or Hogg and Jennings (2001) are coming closer to designing realistic socially rational agents. But, paraphrasing Lissak and Richardson, ‘the limitations of such models must be remembered’. Beyond the scope of this paper, the same type of warning must be addressed to social network theory and its extensive use of graph theory to explain social interactions (Borgatti and Foster 2003). Graphs are merely symbols, in a Peircean sense, displaying some sort of likeness with reality, they are not reality.

Finally, we have to agree with Cederman (2005) on the potential usefulness of Multi Agent Systems for social scientists. But we have to be aware of the fact that the symbolico-cognitive paradigm inherently limits the range of generative ontologies to be created. We have to take even more cautiously his reference to Simmelian social patterns. These patterns are subjectively construed by the observer, they are not the objective reality. Using MABS to replicate these patterns imposes conditions on the artificial agents that need to be socially validated as far as traditional scientific validation is no longer relevant in the case of complex human ecosystems.