Scientists developing Multi-Agent Systems, as part of Distributed Artificial Intelligence (DAI), tend to focus on individual components interacting within a given system (Gilbert and Troitzsch 1999). This is a purely bottom-up approach where representations of the individual components, the agents, display a large autonomy of action. Hence, system-level behaviours and patterns emerge from a multitude of local interactions. Intentionality is deliberately placed at the level of the agents to the detriment of the system itself, greatly limiting its ability to control its own evolution. In the case of human ecosystems, agents can represent individual actors or relevant social groups and communities (Bousquet and LePage 2004). The following definition of a Multi-Agent System (MAS) is generally admitted. A MAS is a conceptual model of an observed system that includes:

An agent is a physical or virtual entity that demonstrates the following abilities: autonomy, communication, limited perception, bounded rationality, and decision-making process based on satisfying goals and incoming information (Ferber 1999). A Multi-Agent Based Simulation (MABS) is the result of the implementation of an operational model (computer-based), designed from a MAS-based conceptual representation of an observed system. The strength of MAS approaches consists in their ability to represent socially and spatially distributed problems. Meaningful examples of application are to be found in ecology (Janssen 2002), sociology (Conte and Castelfranchi 1995), or economics (Tesfatsion 2001).

Cederman (2005) asserts that generative process theorists in social science, shifting from traditional nomothetic to generative explanations of social forms and from variable-based to configurative ontologies, may find in Multi-Agent Systems relevant tools to explore the emergence of social forms in the Simmelian tradition, thanks to common foundations in both epistemology and ontology.

In the following sections of this chapter, we try to evaluate Cederman’s assertion against evidence. First, we describe general features of cognitive agents as stated and used in Artificial Intelligence (AI) research. Then, we argue that our understanding of mental processes, from Bacon’s idols to Tversky’s prospect theory, is inherently limited. In the third section, we question the supposed objective autonomy of agents, drawing from Peirce’s icons and Varela’s enactive cognitive theory. Finally, we propose a way forward that encapsulates the designer and the modelling process into the observed system itself.