Real-time contextual affect-detection from open-ended text-based dialogue is challenging but essential for the building of effective intelligent user interfaces. In our previous work, an affect-detection component was developed, which was embedded in an intelligent agent interacting with human-controlled characters under the improvisation of loose scenarios. The affect-detection module is capable of detecting 25 basic and complex emotions based on the analysis of pure individual turn-taking input without any contextual inference. In this article, we report developments on equipping the intelligent agent with the abilities of interpreting dynamic inter-relationships between improvisational human-controlled characters and performing contextual affect-sensing, based on the discussion topics, the improvisational “mood” that one has created, relationship interpretation between characters, and the most recent affect profiles of other characters. Evaluation results on the updated affect-detection component are also reported. Overall, the performances of the contextual affect-sensing and dynamic relationship interpretation are promising. The work contributes to the journal themes on affective computing, human-robots/agent interaction, and narrative-based interactive theatre.