Can AI build better AI? Qwen-Planner-Agent tests a closed-loop framework on mobile planning

Qwen-Planner-Agent lets AI participate in building the next version of itself. The paper landed on arXiv 24 Sep 2026.

The problem this framework solves

Mobile planning is a demanding test. Complex long-horizon tasks challenge agent reliability. Costly real-device interaction limits development scalability. The framework connects data production model training and deployment through a shared action-feedback-verification contract. This contract is the backbone. It lets each stage speak the same language. Without it the loops cannot close. The contract makes the loop possible.

How the three loops connect

The first loop is AI for Data. Specialized agents construct tasks collect interaction trajectories curate and balance training data and use training feedback to guide subsequent data generation. A human gate sits on this flywheel. The loop turns raw interaction into curated signal.

The second loop is AI for Training. It combines a supervised planning cold start with hybrid-environment online agentic reinforcement learning. The team introduced Competence-Aware Reward-and-Advantage Engineering or CARE to reduce reasoning and tool-use costs while preserving task performance. CARE shapes the reward so the model learns when to think and when to act.

The third loop drives model-harness co-evolution. An execution-evidence-driven loop orchestrates memory skills and tools at runtime and feeds structured action feedback and preserved failure traces back into coordinated model and harness adaptation. The harness evolves alongside the model. Failure traces are not discarded they are preserved and fed back.

Benchmark results

The system achieves the best overall performance among all evaluated models and systems on MobilePA-Bench. It improves over its base model across tool use memory skills and sub-agent coordination. Further evaluations show improvements across non-mobile agentic benchmarks while largely preserving general capabilities. The gains transfer beyond the mobile domain.

What remains unknown

The paper does not say how many real devices were used or what the human gate approval rate looks like. It does not report latency or battery impact on the phone. The cost of the hybrid-environment RL loop in compute hours is not disclosed. The framework remains a research prototype without a public release date.