EcoAgent: An Efficient Device-Cloud Collaborative Multi-Agent Framework for Mobile Automation

TL;DR

EcoAgent addresses the latency and privacy limitations of cloud-centric mobile agents by introducing a closed-loop device–cloud framework with a cloud Planning Agent, a device Execution Agent, and a device Observation Agent. It contributes the Dual-ReACT reasoning paradigm for cloud-based planning with device feedback, and a Pre-Understanding module to compress screen content into textual summaries for privacy and efficiency. Experimental results on the AndroidWorld benchmark show EcoAgent achieves task-success rates competitive with fully cloud-based systems while substantially reducing cloud dependence, token usage, and latency. The work advances privacy-preserving, efficient mobile automation and lays groundwork for practical edge-based deployment as mobile AI hardware advances.

Abstract

To tackle increasingly complex tasks, recent research on mobile agents has shifted towards multi-agent collaboration. Current mobile multi-agent systems are primarily deployed in the cloud, leading to high latency and operational costs. A straightforward idea is to deploy a device-cloud collaborative multi-agent system, which is nontrivial, as directly extending existing systems introduces new challenges: (1) reliance on cloud-side verification requires uploading mobile screenshots, compromising user privacy; and (2) open-loop cooperation lacking device-to-cloud feedback, underutilizing device resources and increasing latency. To overcome these limitations, we propose EcoAgent, a closed-loop device-cloud collaborative multi-agent framework designed for privacy-aware, efficient, and responsive mobile automation. EcoAgent integrates a novel reasoning approach, Dual-ReACT, into the cloud-based Planning Agent, fully exploiting cloud reasoning to compensate for limited on-device capacity, thereby enabling device-side verification and lightweight feedback. Furthermore, the device-based Observation Agent leverages a Pre-understanding Module to summarize screen content into concise textual descriptions, significantly reducing token usage and device-cloud communication overhead while preserving privacy. Experiments on AndroidWorld demonstrate that EcoAgent matches the task success rates of fully cloud-based agents, while reducing resource consumption and response latency. Our project is available here: https://github.com/Yi-Biao/EcoAgent.

Figures (4)

• Figure 1: Comparison of four mobile agent architectures. Cloud-based General Agents exhibit strong planning abilities but suffer from high latency and rely on extra components for visual grounding. Device-based Specialized Agents achieve low latency and strong grounding but cannot plan or reflect. Open-loop Device-cloud Cooperation Agents enable grounding with precise control but lack device-to-cloud feedback. Closed-loop Device-cloud Cooperation Agents (EcoAgent) strike a balance by combining planning, feedback, and grounding, achieving moderate latency and improved adaptability.
• Figure 2: An overview of the EcoAgent framework. The Cloud-Based Planning Agent generates a task plan, which is executed and verified step-by-step by on-device agents. Feedback and pre-understood screen text from the Observation Agent enable dynamic replanning, forming a closed-loop system for efficient and adaptive task execution.
• Figure 3: Illustration of the EcoAgent workflow. The cloud-based Planning Agent is responsible for Dual-ReACT initial planning as well as Memory and Reflection empowered replanning. The device-based Execution Agent and Observation Agent handle action execution and verification, respectively. Together, these three agents collaborate to realize a closed-loop workflow.
• Figure 4: A complete case of EcoAgent executing the ContactsAddContact task in AndroidWorld. The figure illustrates the full interaction loop among Planning, Execution, and Observation Agents across multiple steps until task completion.