The rapid progress of large language models is pushing AI from passive content generation toward active workflows in engineering and scientific discovery. A fundamental question arises: can AI be both the object of development and an active participant in building the next generation of AI systems? We answer this by implementing Qwen‑Planner‑Agent within a closed‑loop AI‑for‑AI framework that supports scalable development and iterative improvement.
Mobile planning serves as a demanding testbed: long‑horizon, complex tasks stress agent reliability, while costly real‑device interaction limits development scalability. The framework links data production, model training, and deployment through a shared action‑feedback‑verification contract.
(i) AI for Data creates a human‑gated data flywheel where specialized agents generate tasks, collect interaction trajectories, curate and balance training data, and use training feedback to steer subsequent data generation.
(ii) AI for Training combines a supervised planning cold start with hybrid‑environment online reinforcement learning. We introduce Competence‑Aware Reward‑and‑Advantage Engineering (CARE) to cut reasoning and tool‑use costs while preserving task performance.
(iii) AI‑driven model‑harness co‑evolution employs an execution‑evidence loop that orchestrates memory, skills, and tools at runtime, feeding structured action feedback and preserved failure traces back into coordinated model and harness adaptation.
On MobilePA‑Bench, Qwen‑Planner‑Agent achieves the best overall performance among all evaluated models and systems, improving tool use, memory handling, skill invocation, and sub‑agent coordination. Additional evaluations show gains on non‑mobile agentic benchmarks while largely retaining general capabilities.
Review: This work demonstrates the feasibility of a closed‑loop AI‑for‑AI ecosystem in high‑cost real‑world settings, offering a systematic path for developing robust mobile planners and other long‑horizon agents.