Research map / Multi-agent systems and games / LLM agents

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

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EcoAgent, a closed-loop device-cloud multi-agent framework where a cloud Planning Agent (Dual-ReACT) plans, on-device Execution and Observation agents act and verify, and a Pre-Understanding Module turns screenshots into short text to cut privacy risk and token cost on AndroidWorld.

Problem

Cloud-based mobile multi-agent systems incur high latency, cost, and privacy risk from uploading screenshots, while existing device-cloud collaboration is open-loop with no device-to-cloud feedback.

Approach

Dual-ReACT (global then local ReACT planning in the cloud producing step-expectation pairs) lets lightweight on-device Observation Agent verify each step against expectations and send compact textual screen summaries back to the cloud, closing the loop.

Key results

• EcoAgent (OS-Atlas) reaches 27.6% SR on AndroidWorld vs 28.4% for cloud-based M3A (two GPT-4o), with 89% fewer MLLM calls and 96% fewer tokens than M3A.
• EcoAgent (ShowUI) achieves 3.9 s per-step latency, versus 15.3 s for M3A and 7.1 s for AppAgent.
• EcoAgent uses 120 kB uplink per task, 17.5x less than AppAgent and 48.6x less than M3A.

Surprising findings

• V-Droid (59.5% SR, device-based) substantially outperforms EcoAgent, which is lower than several cloud and open-loop baselines.
• Adding the Observation Agent with Pre-Understanding raised SR from 15.5% to 25.6% (ShowUI), while only modestly increasing MLLM calls and tokens.

Limitations

• Success rate (25.6-27.6%) is well below the best baselines such as V-Droid (59.5%) and Agent S2 (54.3%).
• Failures dominated by visual grounding errors from small on-device models (10 of 20 sampled failures).
• Device-side models were run on an RTX 3090 server to simulate mobile inference, not on real phones.

Experimental setup

Benchmarks: AndroidWorld

Metrics: success rate average MLLM calls average MLLM tokens latency uplink data volume

Baselines: ShowUI OS-Atlas InfiGUIAgent V-Droid AppAgent M3A Agent S2 UGround-V1-2B UGround-V1-7B MobileAgent CogAgent AutoDroid

Builds on: ReAct M3A UGround Mobile-Agent

Closest papers by meaning

• CORE: Reducing UI Exposure in Mobile Agents via Collaboration Between Cloud and Local LLMs (NeurIPS 2025 · similarity 0.86)
• Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation (AAAI 2026 · similarity 0.85)
• AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents (ICLR 2025 · similarity 0.85)
• SPA-BENCH: A COMPREHENSIVE BENCHMARK FOR SMARTPHONE AGENT EVALUATION (ICLR 2025 · similarity 0.84)
• ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.84)
• VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics (ICML 2026 · similarity 0.83)