This paper focuses on the human-machine communication within the framework of intelligent agents. We propose a generic architecture provided with a natural language (NL) algorithm for command interpretation that can be adapted to different agent's domains. Our NL architecture only depends on the agent's code and its domain ontology. We consider two classical approaches for NL command interpretation: the top-down approach, which relies on the agent's model syntactical constraints, and the bottom-up approach which relies on the set of the agent's possible actions. We propose to combine both approaches in a bottom-up based algorithm that makes use of agent's constraints. We propose a comparative evaluation of these three algorithms.