Conversational agents are seen as a way to offer an anxiety-free L2 practice environment for learners, in the hope that it would produce similar proficiency improvements as those obtained in a chat with a human interlocutor. Various approaches have been used to develop conversational agents: pattern matching (from a knowledge base obtained either manually, from a corpus or with crowdsourcing), AI planners and script-based frameworks. However current agents still present important shortcomings for applications in language learning and very few researchers have ever evaluated their impact on L2 development. Following Petersen’s (2010) successful experiment, we will present a design for a task-based agent that would provide semantically and pragmatically consistent answers inside a specific goal-oriented conversation with a learner.