Qwen-Planner-Agent Lets AI Build the Next Mobile Agent
Can AI be both the thing being built and one of the builders? Qwen's answer, in Qwen-Planner-Agent (arXiv 2609.29892), is a closed loop where agents produce the data, shape the training and adjust the harness for the next version of a mobile planning agent. Mobile planning is a hard proving ground on purpose: long-horizon tasks, and every real-device interaction is slow and costly.
The loop has three parts tied together by a shared action, feedback and verification contract. AI for Data is a human-gated data flywheel where specialized agents construct tasks, collect trajectories, curate and rebalance the training set, and use training feedback to decide what to generate next. AI for Training runs a supervised cold start, then hybrid-environment online agentic RL with a new reward scheme called CARE that cuts reasoning and tool-use cost while keeping performance. The third piece co-evolves model and harness: an execution-evidence loop orchestrates memory, skills and tools at runtime and feeds structured failure traces back into both.
The result is the best overall score on MobilePA-Bench among all evaluated models and systems, gains over the base model in tool use, memory, skills and sub-agent coordination, and improvements on non-mobile agent benchmarks with general capability largely preserved.
The phrase that matters is model and harness co-evolution. Most labs still treat the harness as a wrapper around a trained model. Here failure traces flow into both at once, which is the production version of the argument that the harness is half the product. Note the human gate on the data flywheel. The AI-for-AI loop in a shipping product still keeps a person at the door where data comes in.
← Back to all articles
The loop has three parts tied together by a shared action, feedback and verification contract. AI for Data is a human-gated data flywheel where specialized agents construct tasks, collect trajectories, curate and rebalance the training set, and use training feedback to decide what to generate next. AI for Training runs a supervised cold start, then hybrid-environment online agentic RL with a new reward scheme called CARE that cuts reasoning and tool-use cost while keeping performance. The third piece co-evolves model and harness: an execution-evidence loop orchestrates memory, skills and tools at runtime and feeds structured failure traces back into both.
The result is the best overall score on MobilePA-Bench among all evaluated models and systems, gains over the base model in tool use, memory, skills and sub-agent coordination, and improvements on non-mobile agent benchmarks with general capability largely preserved.
The phrase that matters is model and harness co-evolution. Most labs still treat the harness as a wrapper around a trained model. Here failure traces flow into both at once, which is the production version of the argument that the harness is half the product. Note the human gate on the data flywheel. The AI-for-AI loop in a shipping product still keeps a person at the door where data comes in.
Comments