IT之家 10 月 25 日消息,在 10 月 22 日的 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构建个人 AI 智能体的工艺实践过程。他认为,真正的个人智能体阶段,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,围绕使用者目标自主规划与执行,并进一步发现使用者尚未注意到或有所疏忽的需要。
他表示,个人 AI 智能体应对日常工作、学习、生活难题的理论上限接近 100%,释放这一潜力的关键,在于让平台能力与难题应对路径更好地匹配,借助模型、记忆与生态能力的高质协同,为不同难题找到最优应对路径。
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 多个头部行业的使用者需要透视图,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,持续提高模型理解真实使用者需要、完成复杂任务的能力。
在记忆层面,蒋冠军表示,个人 AI 智能体需要长期理解使用者,识别历史信息与当前需要的关联,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,为其孩子的就业难题给出更贴合家庭处境的建议。
在生态层面,蒋冠军认为,随着模型能力和 Harness 工艺的持续发展,生态的连接与协作能力将成为个人 AI 智能体发展的关键。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,提高复杂现实任务的执行成效。
蒋冠军表示,上述三要素的协同最终指向主动服务,即“记忆给出个人背景,模型结合专业知识识别潜在需要、规划行动,生态支撑执行”。让使用者无需编程,只要说出需要,就能从容搭建用于生活、工作和学习等情形的应用。
本文来源:https://www.ithome.com/1/017/669.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, October 25 — On October 22 at the 2026 China Computer Congress, Jiang Guanjun, AI algorithm lead at Alibaba’s Qwen business unit, shared Qwen’s approach to building personal AI agents. He argued that the true personal-agent stage should deliver incremental value creation and personalized, proactive services, with models, memory, and ecosystem being the three elements that together build this core capability.
Jiang Guanjun divides the development of AI applications into three stages: dialogue, agents, and assistants. Currently, AI applications are accelerating toward the assistant stage. Beyond delivering specific tasks, they are beginning to integrate professional expertise and personal context, autonomously planning and executing around users' goals, and further identifying needs that users may not have noticed or may have overlooked.
He said that the theoretical ceiling for how well personal AI agents can solve the challenges of everyday work, study, and life approaches 100%. Unlocking this potential depends on better aligning platform capabilities with the pathways needed to tackle different problems, leveraging high-quality synergy among the model, memory, and ecosystem capabilities to identify the optimal approach for each challenge.
At the model level, Qwen places a greater emphasis on the stability of task execution. To achieve this, Qwen has built a comprehensive profile of user needs covering over 30 leading industries, identified over 8,000 atomic model capabilities, and engaged industry experts in model evaluation and optimization. This continuously enhances the model's ability to understand real user needs and complete complex tasks.
At the memory level, Jiang Guanjun noted that personal AI agents need to build a long-term understanding of users, identify how past information relates to current needs, and avoid the noise and computational overhead that come from simply accumulating conversations. For example, an AI could take into account a pension-related issue an elderly user mentioned in past conversations and then offer advice on their child’s employment challenge that is better tailored to the family’s situation.
At the ecosystem level, Jiang Guanjun believes that as model capabilities and Harness craftsmanship continue to evolve, ecosystem connectivity and collaboration will become key to the development of personal AI agents. Qianwen is building a model-centric open service system that integrates professional knowledge, tools, industry skills, and third-party agents, connecting services such as e-commerce and local lifestyle offerings, as well as professional capabilities in fields including healthcare and education, to improve the effectiveness of executing complex real-world tasks.
[Name] stated that the synergy among the three aforementioned elements ultimately enables proactive service: “memory provides the personal context; the model combines professional expertise to identify latent needs and plan actions; the ecosystem supports execution.” This allows users to build, with ease, applications for everyday life, work, and learning simply by stating their needs—without any programming.
Source: https://www.ithome.com/1/017/669.htm If there are any copyright issues, please contact us via email at 13951431913#139.com or submit a comment on the right side.