A Case for Declarative LLM-friendly Interfaces for Improved Efficiency of Computer-Use Agents