Incorporating intelligence and social behaviours into virtual worlds for learning is becoming more desirable in making them smart, adaptive, personalized, and therefore, more effective and engaging. Realistic non-player controlled characters (NPCs) are essential of a game world and are making the virtual world more real for players. This is true in video games where more interactive NPCs support the story narrative of a game, making the game more immersive, more convincing, but it is also true in other areas where virtual worlds are used such as business and education, increasing the effectiveness of those environments. Research that has been done with virtual agents and multi-agent systems can be leveraged to create more realistic NPCs through purposeful communication channels, inter-agent interactions and environment-agent interactions for game-based learning applications. This research proposes an approach to controlling avatars with intelligent agents through the creation of an interface between a multi-agent system to a virtual world engine. Basic NPC behaviours controlled by agents using Jason Agent Speak are used to test the feasibility of the approach. A Quizmaster is designed and illustrated as a proof of concept of the use of such agent controlled avatars in educational context.