Qwen outlines personal agent framework: 8,000+ atomic capability requirements, tiered memory, open service ecosystem

Alibaba's Qwen team says a personal agent depends on three parts working together: the model, memory and an ecosystem of services. Jiang Guanjun, head of AI algorithms at Alibaba's Qwen business unit, presented the framework on October 22 at the 2026 China National Computer Congress (CNCC) in Chengdu. The account comes from Chinese-language coverage by Leiphone, Sina Tech and InfoQ China, as reported in English by Pandaily.

Three stages: chat, agent, assistant

Jiang divided AI application development into three stages. In the chat stage, AI answers questions. In the agent stage, it completes tasks it is given. In the assistant stage, which he said the industry is now moving into, AI combines professional knowledge with a user's personal context, plans and acts around the user's goals, and spots needs the user has not noticed. A real personal agent, he said, should create added value and offer proactive, personalized service.

He argued that, in theory, a personal agent could solve close to 100% of everyday problems in work, study and life. The bottleneck, in his account, is matching what the system can do to the right path for each problem.

The three components

Model. Qwen says it focuses on stable task execution. The team built what it calls a user-needs map covering more than 30 major industries and broke it down into more than 8,000 atomic model capability requirements. Industry experts were brought in to evaluate and tune the model against those requirements.

Memory. Jiang said a personal agent has to understand a user over the long term without simply piling up chat history, which adds noise and inference cost. Qwen designs its memory around how strongly events relate to each other, and stores and retrieves them in tiers. His example: an assistant that remembers an elderly user once mentioning their pension could give more fitting advice when the same family later asks about a child's job search.

Ecosystem. Jiang said that as models and harness techniques improve, the ability to connect and coordinate outside services becomes decisive. Qwen is building a model-centered open service system that brings in professional knowledge, tools, industry skills and third-party agents. It links to e-commerce, local services, healthcare and education.

How the parts combine

In Jiang's account, the three parts add up to proactive service. Memory supplies personal background, the model uses expert knowledge to identify needs and plan, and the ecosystem carries out the work. His examples were tracking a patient's rehabilitation schedule and flagging new risks in a user's finances.

Pandaily describes the talk as the most detailed public account yet of the technical thinking behind Qwen Go!, the personal agent Alibaba opened for sign-ups this month.

No model release, benchmark scores, context window or pricing were announced in the talk as reported.

What this means

The talk frames the personal agent as a systems problem rather than a model-quality problem. Qwen's stated approach puts three design questions at the center: how reliably the model executes tasks, how memory is kept useful without the noise and inference cost of raw chat history, and how many outside services the agent can coordinate.

The 8,000-plus atomic capability requirements are Qwen's own figure, and no evaluation results were given alongside it. The same applies to the claim that a personal agent could solve close to 100% of everyday problems, which Jiang framed as a theoretical possibility. Whether Qwen Go! delivers the proactive behavior described will be judged on how it performs for users, not on the framework alone.