The deployment of Artificial Intelligence (AI) on resource-constrained embedded systems, commonly known as Edge AI, is often hindered by stringent memory, computational power, and energy constraints. While the STM32 ecosystem offers specialized tools such as STM32CubeMX and X-CUBE-AI, orchestrating the full Machine Learning Operations (MLOps) lifecycle, from firmware configuration to model optimization and deployment, remains a complex and primarily manual process. This thesis introduces a scalable Agentic MLOps Orchestration Framework designed to automate and optimize the Edge AI development workflow on STM32 microcontrollers. To move beyond the limitations of local execution, the proposed architecture adopts a cloud-native approach that maximizes resource utilization by decoupling the multi-agent orchestration logic (CPU-bound) from the Large Language Model (LLM) inference engine (GPU-bound).