Most agent research focuses on model improvements in isolation; this work argues the bottleneck for robust, real-device mobile agents is the whole development loop — data, training, and runtime harness — not just model size. The core insight: use AI to produce and curate interaction data, adapt training via agentic RL, and iteratively refine a Harness that supplies skills, memory, and execution feedback so model and runtime co-evolve toward reliable multi-app mobile planning.
Key Findings
- Closed-loop AI-for-AI lifecycle: combines AI-assisted task construction and automated trajectory curation (AI for Data), a planning-focused cold start plus hybrid-environment online agentic RL with Competence-Aware Reward-and-Advantage Engineering (CARE) (AI for Training), and an execution-evidence-driven Harness that provides Skills, persistent Memory, and feedback for coordinated updates (AI for Harness). This design aims to turn execution failures into targeted data and harness updates.
- Empirical results: the Qwen-Planner-Agent (27B) achieves an Overall score of 77.05% on MobilePA-Bench, ranking first among evaluated systems. It reports the highest Tool Use score (77.79%) and strong Memory (74.76%) and Skills (86.25%) metrics, while estimating a low output-token cost (~$2.41 per 1,000 tasks under evaluated pricing assumptions).
- Generalization: the Planner Model retains broad agentic capabilities beyond mobile tasks and shows gains on non-mobile agentic benchmarks, suggesting the framework scales to other execution environments.
Who it fits / Tradeoffs
Great fit if you need a practical blueprint for building deployed, multi-app mobile agents where real-device interaction and long-horizon task reliability matter — especially for teams that can integrate a Harness (skills, memory, execution logging) and iterate on data and training. Look elsewhere if you only need a standalone model snapshot (no runtime orchestration) or cannot accept the engineering overhead of maintaining a runtime Harness, hybrid testbeds, and the human gating recommended in parts of the data loop.
Where it fits
This paper sits between agent-model research and systems engineering: it’s valuable as a reproducible development pattern for productizing agents (mobile or otherwise) rather than a purely algorithmic advance. Expect its main returns when combined with real or simulated execution traces and a tool/skill ecosystem that the Harness can manage.