Alibaba Qwen details framework for proactive personal AI agents

From chat to proactive assistance

Alibaba's Qwen team recently outlined a specific roadmap for the evolution of artificial intelligence. Jiang Guanjun, the head of AI algorithms at the Qwen business unit, presented this framework at the 2026 China National Computer Congress in Chengdu. He argued that the industry is currently transitioning from simple chat interfaces to a more complex assistant model.

In the first stage, known as chat, the AI primarily functions by answering questions provided by a user. The second stage is the agent phase, where the system is given a specific task to complete. The third and most advanced stage is the assistant, which the team believes is the current frontier of development.

A true assistant does not wait for every single instruction. It combines professional knowledge with a user's personal context to plan actions around specific goals. According to Jiang, these assistants should be able to spot needs that a user might not have noticed yet, offering proactive and personalized service.

Mapping 8,000 specific capabilities

To make these assistants functional across different parts of life, Qwen is focusing on stable task execution. The team developed a user-needs map that covers more than 30 major industries. This map serves as a blueprint for what the AI needs to know and do in various professional and personal settings.

This map was further broken down into more than 8,000 atomic model capability requirements. These are the smallest building blocks of functionality that the model must master to be useful. To ensure the model meets these standards, the team brought in industry experts to evaluate and tune the AI against these specific requirements.

By focusing on these granular capabilities, the team aims to solve the bottleneck of matching a system's power to the correct path for a specific problem. This structured approach helps the AI understand the nuances of different fields, from healthcare to local services. It ensures the model behaves reliably when it moves from a general conversation to a specific, high-stakes task.

Tiered memory over chat history

One of the biggest hurdles for personal agents is how they remember information over long periods. Jiang noted that simply piling up every chat history creates noise and increases the cost of inference. If an AI remembers every trivial detail equally, it becomes less efficient at finding the information that actually matters for the user.

To solve this, Qwen designs its memory around how strongly different events relate to one another. The system stores and retrieves information in tiers based on relevance. This allows the AI to maintain a coherent understanding of a user's life without getting bogged down by irrelevant data points.

For example, the system might remember a specific mention of a pension from an elderly user. When that same family later asks about a child's job search, the AI can use that relevant background to provide more fitting advice. This tiered approach helps the assistant maintain a long-term understanding of the user's unique situation and preferences.

Connecting to the service ecosystem

The final piece of the framework is a model-centered open service system. As models and technical techniques improve, the ability to coordinate with outside services becomes a deciding factor for success. Qwen is building a system that can link to third-party agents, professional tools, and various industry skills.

This ecosystem connects the AI to practical areas of life such as e-commerce, education, and healthcare. By linking to these services, the AI can move from just planning a task to actually carrying it out. It creates a bridge between the digital intelligence of the model and the physical or digital services people use every day.

The combination of these three parts—memory, model, and ecosystem—enables proactive service. Memory provides the background, the model identifies the needs and creates a plan, and the ecosystem executes the work. This could include tracking a patient's rehabilitation schedule or flagging new risks in a user's personal finances.

The path toward Qwen Go!

This presentation provided the most detailed public account of the technical thinking behind Qwen Go!. This is the personal agent that Alibaba recently opened for sign-ups. The framework explains the underlying logic of how the company intends to turn a large language model into a functional life companion.

Jiang Guanjun suggested that, in theory, a personal agent could eventually solve close to 100% of everyday problems in work, study, and life. While that is a massive goal, the current focus is on creating added value through personalization. The goal is to move away from a tool that requires constant prompting toward one that anticipates the user's next move.

As the technology matures, the focus will likely shift toward refining these 8,000 requirements and expanding the service ecosystem. The success of this framework depends on how well the AI can balance personal privacy with the need to access deep personal context. Users will need to see how these memory tiers and service links operate in real-world scenarios.