10 月 22 日，在 2026 中国计算机大会上，阿里巴巴千问事业部 AI 算法负责人蒋冠军分享了千问打造 Personal Agent 的技术实践。他认为，真正的 Personal Agent 阶段，应该带来增量价值的创造和个性化的主动服务，模型、记忆与生态是共同构建这一核心能力的三要素。
蒋冠军将 AI 应用的发展分为对话、代理和助理三个阶段。当前，AI 应用正加速迈向助理阶段，在交付具体任务的基础上，开始结合专业经验和个人上下文，围绕用户目标自主规划与执行，并进一步发现用户尚未注意到或有所疏忽的需求。
他表示，Personal Agent 解决日常工作、学习、生活问题的理论上限接近 100%，释放这一潜力的关键，在于让系统能力与问题解决路径更好地匹配，通过模型、记忆与生态能力的高效协同，为不同问题找到最优解决路径。
在模型层面，千问更加重视任务执行的稳定性。为此，千问建立了一张覆盖 30 多个头部行业的用户需求透视图，梳理出 8000 多项原子模型能力要求，并引入行业专家参与模型评估与优化，持续提升模型理解真实用户需求、完成复杂任务的能力。
在记忆层面，蒋冠军表示，Personal Agent 需要长期理解用户，识别历史信息与当前需求的关联，避免简单积累对话带来的噪声和推理成本。例如，AI 可以结合老人历史对话提及的养老金情况，为其孩子的就业问题提供更贴合家庭处境的建议。千问通过分析事件关联关系设计记忆体系，并按关联强弱分层存储和调用，提升记忆使用的精准性。
生态层面，蒋冠军认为，随着模型能力和 Harness 技术的持续发展，生态的连接与协作能力将成为 Personal Agent 发展的关键。千问正构建以模型为中心的开放服务体系，整合专业知识、工具、行业 Skill 及第三方 Agent，连接电商、本地生活等服务，以及医疗、教育等领域的专业能力，提升复杂现实任务的执行效果。
三要素的协同最终指向主动服务：记忆提供个人背景，模型结合专业知识识别潜在需求、规划行动，生态支撑执行。例如，在健康管理中持续关注康复安排及其对日常生活的影响，在理财过程中提醒用户留意新的风险与机遇。
在千问上，用户对复杂任务和主动服务的需求持续增长，过去三个月复杂任务平均 Token 消耗量增长 1.6 倍。最新消息显示，千问推出的个人专属 AI 助理「千问 Go!」上线新功能“我的应用”。用户无需编程，只要说出需求，就能轻松搭建用于生活、工作和学习等场景的应用，目前相关功能已开启邀测。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On October 22, at the 2026 China Computer Congress, Alibaba, a thousand times asked the Director of AI algorithms of the Ministry of Industry, Jiang Champion, to share his technical experience in building the Personal Agent. In his view, the real personal Agent phase should bring about the creation of incremental value and personalized proactive services, with models, memory and ecology as the three elements that together build this core capacity.
The Chiang Champion divided the development of AI applications into three phases: dialogue, agency and assistant. Currently, AI applications are accelerating towards the assistant phase, starting to plan and implement autonomously around user targets, based on the delivery of specific tasks, taking into account professional experience and individual context, and further identifying unrecognized or negligent user needs.
The key to unlocking the potential of the Personal Agen, which is close to 100 per cent of the theoretical ceiling for solving day-to-day work, learning and living problems, is to better match the system ' s capacity with the problem ' s path to solving problems and to find optimal solutions to different problems through efficient synergies between modelling, memory and ecological capabilities.
At the modelling level, it is asked that greater attention be paid to the stability of mandate implementation. To this end, a user demand perspective covering more than 30 head industries has been created, more than 800 atomic model capabilities requirements have been streamlined and industry experts have been introduced to participate in model assessment and optimization, and the model's ability to understand real user needs and perform complex tasks has been continuously enhanced.
At the memory level, Chiang Champions stated that Personal Agent needed long-term understanding of users, recognition of the relevance of historical information to current needs and avoidance of the noise and reasoning costs of simple accumulation of dialogue. For example, AI could provide advice on the employment of its children that is more relevant to the situation of the family, taking into account the pension situation mentioned in the historical dialogue of the elderly. The thousand questions will be asked to improve the accuracy of the use of memory by analysing the relationship between events and by designing the memory system and storing and calling it in relation to the strong and weak layers.
At the ecological level, the Chiang champion believes that, with the continued development of modelling capacity and Harness technology, ecological connectivity and collaborative capacity will be key to the development of the Human Alliance. The model-centred system of open services is being built to integrate expertise, tools, industry, Skill and third party Agent, connect services such as electricians, local life, and professional competencies in the areas of health, education, among others, to enhance the effectiveness of the implementation of complex and realistic tasks.
The synergy of the three elements ultimately points to proactive services: memory provides a personal background, models combine expertise to identify potential needs, plan actions, and ecological support for implementation. For example, ongoing attention to rehabilitation arrangements and their impact on everyday life in health management reminds users of new risks and opportunities in financial management.
On a thousand occasions, the demand for complex missions and active services by users has continued to grow, with an average Token consumption of complex missions increasing 1.6 times over the past three months. According to the latest news, the new "My Applications" on-line function of the assistant AI "Manager Go!" has been launched. Users do not have to program, but they can easily build applications for life, work and learning, and the relevant features are now being asked.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.