We present COHIBIT, an edutainment exhibit for theme parks in an ambient intelligence environment. It combines ultimate robustness and simplicity with creativity and fun. The visitors can use instrumented 3D puzzle pieces to assemble a car. The key idea of our edutainment framework is that all actions of a visitor are tracked and commented by two life-like guides. Visitors get the feeling that the anthropomorphic characters observe, follow and understand their actions and provide guidance and motivation for them. Our mixed-reality installation provides a tangible, (via the graspable car pieces), multimodal, (via the coordinated speech, gestures and body language of the virtual character team) and immersive (via the large-size projection of the life-like characters) experience for a single visitor or a group of visitors. The paper describes the context-aware behavior of the virtual guides, the domain modeling and context classification as well as the event recognition in the instrumented environment.