Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

AI summary

Overview

Research area: AI agents for mobile task planning, agentic reinforcement learning, and "AI-for-AI" development pipelines (using AI to build and improve AI systems).

Technical level: Intermediate to Advanced. The paper combines conceptual framework description with formal reinforcement-learning objectives (competence-conditioned rewards, advantage calibration, clipped group-relative policy optimization), so readers benefit from some familiarity with RL and language-model agents.

Scope: The paper reports a framework and system report (arXiv:2609.29892v1, dated 24 Sep 2026, by the MAI Team, Alibaba Token Hub, Alibaba Group) describing how AI is used across data production, model training, and runtime scaffolding to build a mobile planner agent called Qwen-Planner-Agent.

A note on specificity: the paper content available here is truncated and does not report exact benchmark scores, per-capability numbers, dataset sizes, or dollar costs. Where a figure is not stated, this summary says so rather than supplying one. The named numbers that do appear in the available content are the 27B model variant and the CARE hyperparameter example λ_eff = 0.1.

What This Paper Is About

Mobile planning is a hard test case for AI agents: tasks are long-horizon, span multiple applications, require recovering from failure, and need verification that the goal was actually reached, while real-device interaction is expensive and hard to parallelize at scale. The paper asks whether AI can be both the object being developed and an active participant in building the next generation of agent systems, and answers it by constructing Qwen-Planner-Agent inside a closed loop that links data production, model training, and deployed runtime support. The goal is a planner agent that performs reliably on real-world mobile tasks, generalizes to non-mobile agentic benchmarks, and does so at lower estimated output cost than the commercial LLMs in the authors' cost comparison.

Key Contributions

1. A closed-loop AI-for-AI framework. The framework connects data production, model training, and deployment through a shared action–feedback–verification contract. AI turns execution feedback into targeted changes in training data, learning strategies, and runtime support, adapting these decisions as the agent's capabilities evolve and using scalable hybrid environments.

2. Qwen-Planner-Agent, a unified Model–Harness agent system. A trained Planner Model learns task decomposition, grounded tool use, and failure recovery, while a Harness supplies tool-conditioned Skills, persistent Memory, and execution feedback. AI consolidates execution experience and diagnoses failures to guide model training and Harness refinement under human review and versioning.

3. Competence-Aware Reward-and-Advantage Engineering (CARE). Within hybrid-environment online agentic RL, CARE routes trajectory groups to different reward regimes — progress shaping, outcome consolidation, and efficiency refinement — based on group success rate, and calibrates advantages to prevent small efficiency differences from being over-amplified once success saturates.

4. Performance, generalization, and efficiency claims. The system is reported to achieve the highest Overall score among the evaluated frontier models and agent systems on MobilePA-Bench, at a lower estimated per-task output cost (including thinking tokens) than the commercial LLMs in the paper's cost comparison, with the Planner Model also showing competence on general agentic benchmarks.

Main Findings

• Best Overall on MobilePA-Bench: Qwen-Planner-Agent 27B achieves the highest Overall score among the evaluated models and agent systems on MobilePA-Bench. The exact score is not reported in the available content.

• Improvement over the base model: The system improves over its base model across tool use, memory, skills, and sub-agent coordination. Per-capability figures are not reported in the available content.

• Lower estimated output cost: Estimated per-task output cost, including thinking tokens, is lower than that of the commercial LLMs included in the paper's cost comparison. No monetary or token-cost values are reported in the available content.

• Generalization beyond mobile: Evaluations of the Planner Model, Qwen-Planner-Model, show improvements across non-mobile agentic benchmarks while largely preserving general capabilities.

• Model–Harness co-evolution is supported by ablations: Ablation studies support the effectiveness of model–Harness co-evolution in both mobile planning and general agentic settings.

• Efficiency-only advantages are over-amplified by standard normalization: The paper shows analytically that for a fully successful group (all s_i = 1), the GRPO-style standardized advantage reduces to approximately −(e_i − ē)/σ_e, so the efficiency coefficient approximately cancels; even with a small efficiency weight such as λ_eff = 0.1, advantages in success-saturated groups stay approximately unit scale. This is the motivation for the paper's advantage calibration.

• Per-component normalization has a drawback: Normalizing reward components separately before aggregation, as in GDPO, changes each component's effective weight in the original reward space and can alter trajectory rankings under the composed reward — which is why the authors instead preserve the composed reward and put a success-derived floor on the normalization denominator.

Methodology in Plain English

The authors build the agent as two coupled parts — a Planner Model (the policy) and a Harness (the runtime scaffold) — and then organize development as a loop with three stages.

AI for Data is an agent-driven data flywheel. Task-construction agents turn capability requirements into executable task specifications containing a user goal, available resources, initial conditions, target capabilities, and completion criteria, without prescribing a single reference trajectory. Construction deliberately separates scenario coverage (application domains, tools, user intents, interaction patterns) from capability coverage (information acquisition, tool routing, argument grounding, multi-step dependency handling, state tracking, recovery, verified completion). Automated rollouts record trajectories that preserve not just final success but actions, environment feedback, intermediate outcomes, errors, retries, recovery attempts, and completion evidence — failed trajectories are kept because they carry diagnostic value. Curation is separate from verification: verifiers decide whether a task was completed, while the curation workflow decides whether a record is suitable for learning. Low-confidence, conflicting, safety-sensitive, or weakly supported cases are routed to human reviewers, and each release is gated by human approval and versioned. Feedback from rollouts and held-out development-set results drives three updates: down-sampling mastered tasks, up-weighting unstable behaviors, and constructing tasks for missing capabilities.

AI for Training starts with a planning-oriented cold start: supervised fine-tuning on curated data with turn-level error detection, where model-action turns flagged as erroneous during curation are masked out of the loss while verified recovery turns remain supervised. This is followed by hybrid-environment online agentic RL, which runs rollouts in programmatic sandboxes, LLM-simulated environments, and selected real-device sessions. CARE then adapts the reward: a task's group of G sampled trajectories gets a binary success indicator from a rule-based verifier or a rubric-guided generative reward model, and the group success rate selects one of three regimes — progress shaping (below p_low), outcome consolidation (between p_low and p_high), or efficiency refinement (at or above p_high). The scheduling parameters {λ_prog, λ_eff, p_low, p_high} are periodically set by a bounded LLM-based controller from recent training statistics, held-out development-set performance, and a specification of the objective and admissible ranges, and then held fixed for N policy updates. To stop normalization from inflating efficiency-only signals, advantage calibration imposes a floor of sqrt(p_high(1 − p_high)) on the denominator in the efficiency-refinement regime, using the same p_high threshold as reward scheduling. Policy updates maximize a clipped group-relative surrogate objective.

AI for Harness handles deployment as a continuing source of evidence rather than an endpoint. The Harness assembles the model's context from tools, skills, persistent memory, and runtime constraints, and relays structured feedback after each action. AI-assisted diagnosis of preserved traces guides later revision of data, model, and Harness. The paper frames this as maximizing a joint objective over policy parameters and Harness instruction configuration, pursued in practice by alternating model updates (RL under a fixed Harness configuration) with Harness revisions informed by development-set feedback from the updated model. Model and Harness updates are reviewed and versioned offline; model parameters are fixed during serving, and human review is retained for ambiguous and release-critical decisions.

Environment infrastructure: three complementary backends are matched to tasks — programmatic sandboxes for reproducible, verifiable, explicit-state tasks; LLM-simulated environments for long-tail interactions without fixed implementations (with trajectories requiring task-specific validation); and real-device sessions used selectively where live device or service behavior matters. A common agent-facing interface makes these experiences usable in shared data and training workflows, while backend internals stay distinct and simulated and real feedback are not assumed to be equally reliable.

Why This Matters

Impact on research: The paper is a concrete case study of AI participating in its own development loop. Rather than treating data, training, and deployment as separate pipelines, it ties them together with a shared action–feedback–verification contract and lets diagnosed failures propagate back into task generation, training signal design, and runtime scaffolding. The CARE analysis contributes a specific finding: within-group normalization in group-relative RL can erase the intended effect of an efficiency penalty once success saturates, and a success-derived floor on the normalization denominator is one way to keep that dependence intact.

Real-world applications:

• Mobile assistants that carry out multi-step goals across applications, keeping context as state changes and verifying that the intended outcome was reached.
• Enterprise tool-use agents that route structured calls across services and need recovery when a step fails.
• On-device or cloud-served agents where per-task output cost, including thinking tokens, is a practical constraint.
• General agentic workflows outside mobile — the paper reports improvements on non-mobile agentic benchmarks, including structured tool use over services such as Model Context Protocol (MCP) servers.

Industry relevance: The system is developed and reported by the MAI Team at Alibaba Token Hub, Alibaba Group, and the framing emphasizes the economics of agent development: costly, low-parallelism real-device interaction is supplemented by scalable sandbox and simulated environments, and the deployment harness is versioned rather than updated autonomously in production. Cost-aware evaluation relative to commercial LLMs places the work in the practical deployment conversation rather than only the benchmark conversation.

Future Directions

• Quantifying the co-evolution loop. The paper presents alternating model and Harness updates as pursuing a joint objective, but the available content does not report how much each channel contributes separately at scale.
• The gap between simulation and real devices. The framework explicitly does not assume simulated and real-device feedback have identical reliability; determining how far simulated rollouts can substitute for real-device sessions remains an open question.
• Governance of AI-proposed changes. Human review is retained for ambiguous and release-critical decisions, and each data release preserves a versioned history. How much of the loop can be safely automated, and where review must remain, is left open.
• Extending beyond mobile and toward general agentic settings. The paper reports improved non-mobile agentic benchmark results with general capabilities largely preserved; measuring and protecting that preservation as the loop continues is a natural next step.
• Exact reported numbers. The available content does not report the specific Overall score, per-capability MobilePA-Bench results, dataset sizes, or cost figures, so reproducing or comparing against those figures requires the full paper.

Target Audience

This paper is most useful to agent researchers and engineers building long-horizon tool-using agents, to RL practitioners working on multi-turn agentic training and reward design for group-relative methods, and to product and platform teams deciding how to structure data flywheels, hybrid simulation infrastructure, and a deployable agent harness. Readers seeking only final benchmark tables will find the available content short on numbers; readers interested in how a development loop is organized end-to-end, and in a specific pitfall of advantage normalization in group-relative RL, will find it most relevant.

Authors’ abstract

The rapid progression of large language models is extending AI from passive content generation into the active workflows of engineering and scientific discovery. This shift raises a compelling question: can AI be both the object of development and an active participant in building next-generation AI systems? We explore this question by building Qwen-Planner-Agent within a closed-loop AI-for-AI framework for scalable development and iterative improvement. Mobile planning offers a demanding test of this approach: complex, long-horizon tasks challenge agent reliability, while costly real-device interaction limits development scalability. The framework connects data production, model training, and deployment through a shared action-feedback-verification contract. (i) AI for Data builds a human-gated agentic data flywheel in which specialized agents construct tasks, collect interaction trajectories, curate and balance training data, and use training feedback to guide subsequent data generation. (ii) AI for Training combines a supervised planning cold start with hybrid-environment online agentic reinforcement learning, where we introduce Competence-Aware Reward-and-Advantage Engineering (CARE) to reduce reasoning and tool-use costs while preserving task performance. (iii) AI drives model--harness co-evolution through an execution-evidence-driven loop that orchestrates memory, skills, and tools at runtime and feeds structured action feedback and preserved failure traces back into coordinated model and harness adaptation. Qwen-Planner-Agent achieves the best overall performance among all evaluated models and systems on MobilePA-Bench, improving over its base model across tool use, memory, skills, and sub-agent coordination. Further evaluations of our model show improvements across non-mobile agentic benchmarks while largely preserving general capabilities.