In the context of a ECA-human interaction, we have created a BDI-like reasoning engine based on the agent's mental states. This reasoning engine first aims to trigger the agent's emotions from its goals, beliefs, ideals and the notion of responsibility. Then this engine selects the agent's communicative intention from its mental states and a set of dialogue rules. The integration of "Stimulus Evaluation Checks" from Scherer's appraisal theory allows us to associate the selected communicative intention with a multimodal expression. We present a test-scenario involving an argument between the user and the ECA MARC, currently used to evaluate the perceived sincerity and believability of the ECA behaviour.