Tongyi-MAI/Qwen-Planner-Agent

What it claims

The repo hosts the blog and technical report site for Qwen-Planner-Agent — explicitly not the release repo for weights, training code, or the agent implementation (stated in both English and Chinese). Its distinguishing claim beyond the raw score: the 27B model's 77.05% Overall on MobilePA-Bench is 9.83 percentage points above its own baseline, and it ranks first in Overall among the evaluated systems in a comparison spanning tool use, memory, Skills, sub-agent coordination, and estimated output cost (the $2.41 / 1,000 tasks figure).

How it works

The framework is organized as a closed loop across three stages, where execution evidence from real runs feeds back into each:

flowchart LR
  D[AI for Data: task construction,<br/>trajectory collection, failure analysis] --> T[AI for Training: planning SL +<br/>hybrid-env RL with CARE]
  T --> H[AI for Harness: tool-conditioned Skills,<br/>evidence-grounded memory]
  H -->|execution evidence| D

CARE (the page calls it Competence-Aware Reward-and-Advantage Engineering, differing from the summary's "Competence-Adaptive") adapts rewards and advantages to the model's current competence, trading off task completion against reasoning and tool-use efficiency. The Harness is stateful: it retrieves task-relevant evidence, retains constraints across turns, reconciles prior dialogue with tool results, and revises only requested settings. On coordination, the planner respects action prerequisites, hands confirmed artifacts between sub-agents, and revises when tools return errors.

Evidence

Only the single benchmark line is quantified: 77.05% Overall, +9.83pp over baseline, ~$2.41/1,000 tasks (output tokens only, including thinking tokens; excludes input tokens, external tool charges, device execution, and Harness processing). No per-column numbers, dataset sizes, or compute are given in the fetched text.

Limits

There is nothing to run or adopt here — no weights, no code, no install commands. Actionable material is the technical report and recorded example replays (which preserve full tool arguments/returns including failed attempts and corrections), both hosted on the external project website. Cited as a 2026 tech report by "MAI Team, Alibaba Token Hub."