AI research wiki
agentic-systems · 2026-09-26 · Tier 2

Qwen-Planner-Agent: a closed-loop AI-for-AI framework for mobile planner agents

Qwen-Planner-Agent: a closed-loop AI-for-AI framework for mobile planner agents

Source: arXiv 2609.29892 (Qwen team). HuggingFace Daily Papers (09-25, 9 upvotes) and Kurate cs.AI weekly #20, so cross-source confirmed (HF + Kurate). Raw: raw/huggingface/2026-09-25-qwen-planner-agent-a-closed-loop-ai-for-ai-framework-for-rea.md, raw/kurate/2026-09-26-cs-ai.md

TL;DR

A mobile planner agent has to carry out long, multi-step tasks on a phone: open apps, chain tool calls, remember what it did. Building one is slow because real-device interaction is expensive and long-horizon tasks fail in many ways. The Qwen team builds the agent with agents. One shared "action, feedback, verification" contract connects three loops. AI for Data uses specialized agents to create tasks, collect trajectories, curate and rebalance training data, and use training feedback to steer the next round of data, with humans gating the flywheel. AI for Training does a supervised planning cold start and then online agentic RL across hybrid environments, with a new reward scheme, CARE (Competence-Aware Reward-and-Advantage Engineering), that penalizes unnecessary reasoning and tool calls once the model is competent at a task. Model-harness co-evolution feeds structured failure traces from runtime back into both the model and the harness (its memory, skills and tools). The result tops MobilePA-Bench and improves on non-mobile agent benchmarks while largely keeping general capability.

flowchart LR
  D[AI for Data:<br/>agents build tasks,<br/>trajectories, curation] --> T[AI for Training:<br/>SFT cold start +<br/>online agentic RL]
  T --> C{CARE reward:<br/>cut reasoning +<br/>tool cost when<br/>competent}
  C --> M[Model]
  M --> H[Harness:<br/>memory, skills,<br/>tools at runtime]
  H -->|failure traces| D
  H -->|failure traces| T
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class D input
  class C decision
  class M,H output
  class T aux

Key points

  • The efficiency hook is CARE. Reward shaping that scales the cost penalty on reasoning tokens and tool calls with the model's competence on a task. The agent is allowed to think hard on tasks it has not mastered and is pushed to be terse on ones it has. The abstract claims reduced reasoning and tool-use cost with task performance preserved, but gives no numbers.
  • Model and harness are trained as one system. Failure traces are preserved and routed to both, instead of treating the harness as a fixed wrapper around a model.
  • Human-gated flywheel. Agents do the data work; humans approve it.

Gaps

  • No cost numbers in the abstract, and MobilePA-Bench appears to be the team's own benchmark.
  • Kurate's placement carries little weight this week: its three-LLM tournament has not scored any paper for weeks (every entry sits at the 1200 default), so rank #20 is effectively recency.

How this relates to prior wiki pages

  • Joins the self-improvement cluster on self-evolving-agents, and specifically the harness-plus-model co-design line in agent-harness-engineering.
  • CARE is a training-time version of test-time compute allocation (spend reasoning where it is needed), the theme of test-time-compute-allocation. It shares the 09-25 HF list with Rufus-Air (an open eight-stage post-training recipe on GLM-4.5-Air), and both treat the pipeline's structure, not only the model, as the thing being engineered.

Links