Abstract
Cloud-based mobile agents powered by (multimodal) large language models((M)LLMs) offer strong reasoning abilities but suffer from high latency andcost. While fine-tuned (M)SLMs enable edge deployment, they often lose generalcapabilities and struggle with complex tasks. To address this, we propose\textbf{EcoAgent}, an \textbf{E}dge-\textbf{C}loud c\textbf{O}llaborativemulti-agent framework for mobile automation. EcoAgent features a closed-loopcollaboration among a cloud-based Planning Agent and two edge-based agents: theExecution Agent for action execution and the Observation Agent for verifyingoutcomes. The Observation Agent uses a Pre-Understanding Module to compressscreen images into concise text, reducing token usage and communicationoverhead. In case of failure, the Planning Agent retrieves screen historythrough a Memory Module and replans via a Reflection Module. Experiments onAndroidWorld show that EcoAgent achieves task success rates comparable tocloud-based mobile agents while significantly reducing MLLM token consumption,enabling efficient and practical mobile automation.