Qwen-Planner-Agent: a closed-loop AI-for-AI framework for mobile planner agents
Tongyi-MAI · hf · 2026-09-25
Tongyi-MAI built Qwen-Planner-Agent within a closed-loop AI-for-AI framework, exploring whether AI can be both the object of development and an active participant in building next-gen AI systems, tested on demanding long-horizon mobile planning. A shared action-feedback-verification contract connects data production, training, and deployment.
Three loops: (i) AI for Data — a human-gated agentic data flywheel where specialized agents build tasks, collect trajectories, and curate data guided by training feedback; (ii) AI for Training — supervised planning cold start plus hybrid-environment online agentic RL with CARE (Competence-Aware Reward-and-Advantage Engineering) to cut reasoning/tool-use cost while preserving performance; (iii) model–harness co-evolution via an execution-evidence loop orchestrating memory, skills, and tools at runtime.
Qwen-Planner-Agent achieves the best overall score on MobilePA-Bench, beating its base model on tool use, memory, skills, and sub-agent coordination, while improving on non-mobile agentic benchmarks and preserving general capabilities.
More from coding & agent
- Meta launches Muse Code terminal coding agent, resets usage limits for DevDay — armand_ruiz · 2026-09-25
- User ditches GUI: local LLM now his main PC interface for installs, GRUB and more — Forward_Jackfruit813 · 2026-09-25
- Where an agent loop runs, how it fails, and how to build durable sessions — tempNull · 2026-09-25
- Community turns the Opus 5.5 pixel-art prompt into an open-source pixel-anims skill — majidmanzarpour · 2026-09-25
- Opus 5.5 + one prompt generates animated pixel art wizard purely in code — majidmanzarpour · 2026-09-25
- A visual harness for Claude Code: watching your agents work after hours — EricBuess · 2026-09-25