This master’s thesis aims to refine an LLM-based feedback system for program- ming assignments in the Artemis Learning Management System. Despite previous advancements, challenges remain in providing accurate, adaptive, and scalable feedback. The proposed enhancements include the integration of model vector databases, specialized models like Code Llama 2, and an agentic-based approach to address issues of language compatibility and feedback precision. The objective is to reduce manual workload for educators and provide immediate, tailored feed- back to students, thereby enhancing the educational experience and efficiency in learning programming skills.