Game-Based Learning Environments (GBLE) show promise in fostering Computational Thinking (CT) and programming skills in K-12 educations age, by letting students design and run algorithms to advance in games by means of simplified programming blocks. However, such GBLEs can be challenging to some students due to their open-ended nature and minimal guidance. Research has shown that these difficulties can be in part alleviated by adaptive support meant to help students understanding what to do to complete programming problems. In this project we aim to design and evaluate in these GBLE AI-driven support by leveraging a personalization model meant to predict what help content the students need, depending on their current solution, interaction behaviors and attentional eye movements patterns. To this end, we will use a two-phases research approach. First, we will leverage multimodal student data to build AI models able to predict at runtime the CT concepts a student needs scaffolding with, and what level of help is required. This will provide an AI-driven personalization model that can trigger dedicated help interventions. Second, we will investigate how to deliver the help within target GBLEs in an effective and non-intrusive way, by leveraging in particular pedagogical agents that can offer a more engaging and human-like form of interaction. We will use as test-beds two serious games that introduce CT and programming to elementary and high school students, Pyrates and Spy. The main goal of the research is both to provide and evaluate a personalization model for fostering CT concepts, as well as insights on the best way to deliver this help to the students. In the long-run the personalization aim at easing the adoption of GBLEs at school in CT courses.