This paper addresses the issue of hybridization between reactive and cognitive approaches within a single decision-making architecture for virtual agent in an urban simulation. We use a reactive module in order to manage reactive behaviors and agent autonomy, and a cognitive module for anticipation, learning and complex behaviors management. The purpose of the cognitive module is to increase the agent's behavior credibility. The agent's reactive and proactive behaviors are sent to a decision module which is able to integrate, decompose, combine and select an action.