Three agents turn the smartphone into an execution environment

Qwen Intelligence combines three agents specialized in planning, smartphone interaction, and content creation. Alibaba is releasing the architecture alongside a set of open benchmarks covering mobile tasks, cross-app execution, real-device performance, and safety.

One agent plans, another acts

The next environment for AI agents fits in a pocket. With Qwen Intelligence, Alibaba combines three specialized systems that divide the different stages of completing a task on a smartphone.

The Mobile Planner Agent interprets a request, breaks it down, and coordinates the tools or agents required to complete it. The Mobile-Use Agent handles interactions with the phone itself, prioritizing APIs when available and falling back to graphical interface control when an operation requires direct interaction with an app.

The third component, Mobile Creative Agent, focuses on turning short instructions into finished content. Qwen reports a first image generation time of roughly three seconds, around twice as fast as the systems included in its comparison. That figure comes from Alibaba's own evaluation. Planning with tools, memory, and multiple agents

The Mobile Planner Agent does more than generate a sequence of steps. Its architecture combines a Planner Model with a system that maintains task state, retrieves relevant memories, calls tools and Skills, and delegates operations to specialized sub-agents.

Execution remains part of the planning loop. When a tool fails or returns an unexpected result, the planner can revise its approach instead of continuing through a predetermined sequence.

On MobilePA-Bench, Qwen-Planner-Agent 27B reaches a 77.05% Overall score in Alibaba's published results. GPT-6 Astra records 76.84%, Claude Opus 5 75.71%, and Qwen 3.8 Max 71.77% under the same evaluation. The benchmark covers tool use, memory, Skills, and sub-agent coordination. Moving between APIs and the phone screen

The Mobile-Use Agent handles execution. It can search for information, compare options, organize actions, and move across multiple applications by combining tool calls with direct interface interaction.

Alibaba reports an 82.1% success rate on MobileWorld and 92.2% on MobileWorld-Real, which evaluates the agent on physical devices. The current technical page also reports 97.5% on AndroidDaily.

The system extends beyond smartphones. Published evaluations also cover computer and browser use, with 79.5% on OSWorld-Verified and 73.6% on WebArena. These remain results reported by the Qwen team under its evaluation setup. Open benchmarks alongside the agents

Alibaba is also publishing the environments used to evaluate these systems. MobilePA-Bench tests planning inside an interactive, persistent mobile environment rather than relying on a static collection of API calls.

The benchmark contains 1,705 tasks and 212 tools across 13 functional domains. It also evaluates sub-agent collaboration, memory use, and the ability to invoke preconfigured Skills.

MobileWorld focuses more directly on execution and workflows spanning multiple applications. Qwen complements it with MobileWorld-Real for physical devices and MobileWorld-Safety for safety-related behavior.

Qwen Intelligence brings planning, interface control, and content generation into a single mobile architecture, while its individual components remain documented separately through their technical reports and benchmark suites.