Alibaba Tongyi Lab (Tongyi-MAI, @Alibaba_Qwen) has open-sourced Qwen-Planner-Agent (arXiv: 2609.29892, Project: tongyi-mai.github.io/Qwen-Planner-Agent/), a next-generation mobile planner agent operating on Android systems. Overcoming the bottleneck of costly real-device interaction traces and brittle multi-turn reliability, Qwen-Planner-Agent establishes a closed-loop 'AI-for-AI' co-evolution framework. Integrating an agentic data flywheel, Competence-Aware Reward-and-Advantage Engineering (CARE) reinforcement learning, and execution-evidence model adaptation, it achieves top rank on MobilePA-Bench.

Key Takeaways

  • ✓Pioneers a closed-loop AI-for-AI mobile agent lifecycle connecting automated trajectory synthesis, RL, and deployment.
  • ✓Introduces Competence-Aware Reward-and-Advantage Engineering (CARE) to curtail token reasoning costs while boosting tool precision.
  • ✓Orchestrates dynamic model-harness co-evolution, harmonizing runtime memory, skills, and hierarchical sub-agents.
  • ✓Secures top overall performance on the MobilePA-Bench suite, outperforming prior proprietary and open mobile agents.
  • ✓Interactive project website, evaluation suite, and behavioral action traces released openly on GitHub.
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
Qwen-Planner-Agent 官方项目主页与视频
tongyi-mai.github.io/Qwen-Planner-Agent/
↗
🌐
ArXiv 论文预印本 (2609.29892)
arxiv.org/abs/2609.29892
↗
📝
Hugging Face Daily Papers 讨论主页
huggingface.co/papers/2609.29892
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points Mobile operating systems represent a critical frontier for autonomous agents, requiring models to navigate asynchronous UI refreshes, unpredictable popups, and multi-app workflows (e.g., cross-platform travel booking). However, real-device execution cannot be parallelized like game simulators. Gathering human interaction trajectories on physical smartphones is prohibitively slow, while existing agents suffer from context drift and circular tool invocations on extended step horizons. ### Architecture Highlights & Internals Tongyi Lab introduces an AI-for-AI co-evolution framework structured into three interconnected loops: (1) AI for Data deploys specialized agent teams to synthesize scenarios, collect device trajectories, and balance curriculum difficulty based on upstream training feedback; (2) AI for Training leverages Competence-Aware Reward-and-Advantage Engineering (CARE) within online RL, penalizing redundant screen swipes and reasoning bloat; and (3) Runtime Model-Harness Co-Evolution logs structured failure traces into an execution-evidence database to iteratively refine skills, memory schemas, and sub-agent routing. ### Authoritative Benchmarks & Measured Scores Evaluated on the MobilePA-Bench benchmark spanning over 100 diverse Android applications: Qwen-Planner-Agent ranks first across tool precision, memory retention, and multi-agent coordination. CARE reinforcement learning slashed average token consumption by 34.2% while reducing physical interaction steps by 21.6%. General reasoning evaluations confirmed zero degradation on non-mobile tasks. ### Developer Hands-on Guide Mobile system architects and RPA engineers can review action schema specifications and video demonstrations at https://tongyi-mai.github.io/Qwen-Planner-Agent/, and access the preprint on arXiv and Hugging Face Papers.