Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

Qwen-Planner-Agent. A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents. By Tingyu Qu and twenty-five others, from Alibaba Token Hub, Alibaba Group. Published on the twenty-fourth of September. No code repository is linked to it.

AI is moving out of passive content generation and into the active workflows of engineering and scientific discovery, carried by fast progress in large language models. That raises a question the authors put directly: can AI be both the thing being developed and an active participant in building the next generation of AI systems? To test it they need a demanding target, and they choose mobile planning. It is demanding on two fronts. The tasks are complex and long-horizon, which strains an agent's reliability. And interacting with real devices is costly, which limits how far development can scale.

Qwen-Planner-Agent is built inside a closed-loop, AI-for-AI framework meant for scalable development and iterative improvement. The framework ties together three stages into one loop: producing data, training the model, and deploying it. What binds them is a shared contract of action, feedback, and verification, so that each stage speaks the same language to the others. Each stage then has AI doing the work.

The first stage is AI for Data. It is a human-gated agentic data flywheel, in which specialized agents construct tasks, collect interaction trajectories, curate and balance the training data, and use feedback from training to guide what data gets generated next.

The second is AI for Training. A supervised planning cold start is followed by online agentic reinforcement learning in a hybrid environment. Here the authors introduce Competence-Aware Reward-and-Advantage Engineering, or CARE, whose purpose is to cut the cost of reasoning and tool use while preserving task performance. The third stage is where the model and its harness co-evolve. An execution-evidence-driven loop orchestrates memory, skills, and tools at runtime, and feeds structured action feedback and preserved failure traces back into adapting the model and the harness together, in coordination.

On MobilePA-Bench, Qwen-Planner-Agent achieves the best overall performance among all the models and systems evaluated. It improves over its own base model across four areas: tool use, memory, skills, and coordination between sub-agents. The gains are not confined to mobile. Further evaluations show improvements on agentic benchmarks outside the mobile setting, while general capabilities are largely preserved.

The framework answers its opening question with a working system rather than an argument. Agents build the tasks and curate the data. Training feedback steers the next round of data. Runtime failures are kept and sent back to reshape both the model and the harness around it. At every stage, AI is both the thing being improved and one of the hands doing the improving, and the contract of action, feedback, and verification is what lets those stages hand work to one another.