Edge AI integrates AI techniques with heterogeneous mobile devices to enable perception and actuation in real-world environments, facilitating applications such as smart sensing 1 and healthcare 2. To reduce the burden of developing such applications, recent works 3, 4 build agentic systems based on Large Language Models (LLMs) to automatically translate user requirements into executable edge AI programs. However, these systems typically rely on handcrafted, predefined agentic workflows, and therefore often struggle to handle diverse mobile devices, heterogeneous runtime environments, and dynamic resource constraints in real-world scenarios. As a result, even well-engineered agents may struggle to accommodate such variability, suffering from inflexibility, limited adaptability, and high maintenance costs.
Shen et al. (Thu,) studied this question.