The paper aims at presenting an agent-based modeling exercise to illustrate how small differences in the cognitive properties of agents can generate very different macro social properties. We argue that it is not necessary to assume highly complicated cognitive architectures to introduce cognitive properties that matter for computational social science purposes. Our model is based on different simulation settings characterized by a gradual sophistication of behavior of agents, from simple heuristics to macro-micro feedback and other second-order properties. Agents are localized in a spatial interaction context. They have an individual task but are influenced by a collective coordination problem. The simulation results show that agents can generate efficiency at a macro level particularly when socio-cognitive sophistication of their behavior increases.