在1024 模力工场 · AI社区日主论坛上，阿里云智能集团研发副总裁、终端智能计算事业部负责人张献涛，以《云电脑的演进：从人的办公桌面到智能体的工位》为题，分享了云电脑在 Agent 时代的新角色：为智能体提供一个能持续运行、权限可控，并能与人协作的工作环境。
张献涛于 2014 年加入阿里云，长期负责弹性计算服务 ECS，随后转向终端设备与 AI 结合的业务探索。从支撑云上业务运行，到为 Agent 提供操作软件的环境，他关注的问题也随之延伸：当智能体开始进入企业工作流程，怎样才能让它顺利干活？
过去，云电脑是人远程办公的工具；如今，Agent 也需要这样一个“工位”。它要在里面打开软件、访问业务系统、处理任务，遇到无法继续的步骤，还得有人能接手。
给 Agent 配“工位”，首先是让它有工具可用。 之前的 Agent 主要通过浏览器执行任务；接入云电脑后，它们可以进一步使用桌面软件和企业应用，处理跨系统的复杂工作。对于没有开放 API、仍然依赖图形界面操作的老旧系统，这种能力尤其重要。
但有一台电脑，并不意味着 Agent 就能在企业里顺利工作。它以什么身份登录，能访问哪些数据、执行哪些操作，操作过程如何记录和审计，都需要明确。云电脑的作用，也因此从提供运行环境，延伸到身份、权限和资源的统一管理。
另一个关键问题是：Agent 卡住了，人能不能接得上手？
张献涛以订单处理为例：Agent 接收新订单后，需要核对信息、生成报告，再提交主管审批。这条流程看起来顺畅，实际执行时却可能遇到登录失效、验证码，或需要人工判断的情况。任何一步受阻，都可能让任务停下来。
这时，云电脑需要保存当前工作状态，通知用户并说明阻塞原因，让人接入处理。问题解决后，Agent 再接着执行。这样的协作，既需要人看得懂任务进行到了哪一步，也需要系统保留现场，避免每次接管后都从头重来。
这些能力已经开始用于具体业务。据张献涛介绍，好未来的一个项目在不到两周内，通过企业级云电脑部署了约 1,000 个 Agent，为教师提供教案、课件制作及作业批改等支持。员工也可以进入云电脑进行调试，并利用数据恢复等能力保障数据安全。
总而言之，在张献涛看来，云电脑积累了多年的身份权限管理、资源管理和协作能力，刚好可以成为 Agent 的“工位”：云电脑为 Agent 提供运行软件、访问业务系统的环境，并管理好身份和权限。当 Agent 遇到问题时，人可以接入同一台云电脑处理，随后再由 Agent 继续执行任务。
以下为张献涛讲演的完整速记，模力工场在不改变原意的前提下进行了订正和编辑。
Agent 也有“工位”了：云电脑让它能接手业务，也管住了权限
各位嘉宾、各位朋友，大家节日好！非常欢迎大家来到云谷，也感谢 Kevin 的邀请。过去几年，我和 Kevin 有不少交流。今天也很开心见到许多老朋友，刚才还见到了毕玄。
毕玄见面后的第一句话就是：“你们的云电脑现在已经变成 Agent 的基础设施了，应该卖得很好吧？”我说，确实如此。
先向在座的新朋友介绍一下自己。我的履历比较简单：毕业后的前 9 年在英特尔工作，主要从事云计算核心技术相关的研发。2014 年加入阿里云，到现在已经 12 年了。前 10 年，我主要负责弹性计算服务，也就是 ECS。2024 年初，我开始内部创业，探索终端设备与 AI 结合的新业务。
过去几年，我们做了无影云电脑。前段时间的云栖大会上，我们还发布了千问 Book，这是业界第一台智能体电脑。
今天讨论的话题与此密切相关：无影云电脑迎来了一个新时代，它正在成为智能体的“工位”。
首先，回顾过去三年多的变化。2023 年初，以大模型为代表的这一轮 AI 浪潮爆发，大致经历了三个阶段。
第一个阶段，以 ChatGPT 这样的 Chatbot 为代表。它带给大家的震撼是：人工智能的能力确实很强，已经能够轻松处理自然语言，与人类交流。
第二个阶段，以 Agent 应用为代表，比如 Coding Agent 和工作类 Agent。它们让我们发现，AI 可以调用各种工具，代替人类完成非常复杂的工作。
第三个阶段，是最近很受关注的 Personal Agent，也就是个人智能体。各家厂商都在发布相关产品，阿里也发布了千问 Go（千问在 2026 年推出的 Personal Agent）。
这意味着 Agent 正在进入一个非常重要的场景：个人生活和工作。
关注 Personal Agent 的朋友应该能感受到，一个被反复强调的重点是：给每个 Agent 配备一台云电脑，让 Agent 可以向云电脑下达各种复杂任务。以前，它可能主要调用浏览器；有了云电脑之后，就能完成更加复杂的工作。
背后的道理很简单。我们接触的世界，可以分为物理世界和数字世界。物理世界是我们可以触及的现实世界，数字世界则包括互联网和移动互联网，是人类构建起来的。
Agent 如果真正想帮助人类、代理人类完成更加复杂的工作，就需要使用人类使用的工具，与数字世界打交道。
所以，就像刚才毕玄所说，云电脑变成了 Agent 的基础设施。我们平时开玩笑说，这就是给 Agent 安排了一个很好的“工位”。
再来看云电脑的演进过程。
第一代云电脑更多出现在呼叫中心等场景，当时通常叫 VDI。很多企业会在数据中心部署 VDI 或云电脑。
发展到第二个阶段，云电脑已经可以完成更加复杂的工作。以无影云电脑为例，我们已经服务了几十万家企业，每天有数百万用户在云上办公。
这对企业和个人都有很大价值。对企业来说，云电脑有助于更好地管理 IT 资源，打通内部系统。当业务系统、办公系统都在云上时，效率会有很大提升。
对员工来说，无论使用手机、平板，还是通过网页，都可以方便、安全地连接公司的系统。
随着业务发展和负载变化，过去几年，我们还推出了 GPU 工作站。三维设计、仿真，以及 AI 模型的训练和推理，都可以通过这类工作站完成。
随着这些能力不断积累，进入 AGI 时代，云电脑也迎来了新的发展机会。它可以接受 Agent 派发的任务，让 Agent 调用电脑中的各种复杂软件。
比如，大模型可以通过 Computer Use 操作软件，完成视频制作、3D 制作等任务。
在这样的场景下，云电脑变得越来越重要。要满足 Agent 的需求，需要哪些能力？这些恰好也是我们过去多年在云电脑领域积累的能力。
首先，云电脑不只是云上的一台电脑，还提供身份和安全管理，包括认证、审计，以及对各类系统的访问权限管理。
其次，是交互与系统能力。通过自研协议，用户可以完整使用云上的操作系统，在其中安装各种软件，并进行清晰、流畅的交互。
第三，是资源和生命周期管理。企业采购实体电脑，通常需要一定周期；但在云上，只需要几条命令，就可以创建上千台、上万台电脑，并完成整个生命周期的管理。
这些能力如何对应 Agent 时代对电脑的要求？我们做了进一步梳理。
第一，是“工位”。云电脑可以为 Agent 提供一个 7×24 小时安全、可靠、稳定运行的云上环境，随时接受 Agent 的任务需求。
第二，是管控。我们在身份、认证等方面积累的能力，可以让 Agent 接入企业系统时拥有自己的数字身份，也就是 Agent ID，并在相应权限下代理用户完成企业内部的复杂任务。
第三，是协同。Agent 在执行任务时，经常需要 Human-in-the-loop，也就是人的参与。遇到无法独立解决的问题，比如验证码，或者需要登录某个系统时，它仍然需要把控制权交给人。
云电脑已经具备这方面的能力。Agent 通知你之后，你可以通过不同设备接入，与 Agent 协同完成任务。
为什么 Agent 还需要一个持久化的环境？我们用一个例子来说明。
很多制造企业都有 ERP 系统，制单过程对人来说相对简单。假设给 Agent 设置一个定时任务：每天有新订单时，自动完成订单及相关系统信息的校验，生成报告，再推送给主管审批。
这个过程可能遇到各种异常，导致 Agent 无法自动完成任务，需要人工介入。遇到问题时，Agent 如何与人有效互动，也是需要解决的问题。
这些能力，云电脑在过去多年中已经有所积累。有些系统已经可以通过 CLI 操作，有些可以直接调用 API，但也有一些系统，特别是老旧系统，仍然需要通过图形界面操作。
在这种跨系统、跨应用的复杂任务中，人与 Agent 的协同非常重要。如果没有一台好用的云电脑提供支持，Agent 就很难完成这类任务。
此外，Agent 进入企业生产系统或生产环境后，也会面临不少挑战。
比如，让 Agent 每周代理你生成一份周报，过程中同样可能遇到各种异常。复杂任务怎么处理，各个系统之间的权限怎么管理，并不是给它一台电脑就能解决的。
企业还需要统一管理和管控，让 Agent 拥有自己的 ID 和相应权限，才能连接不同系统，完成复杂任务。
Agent 也需要与人交互，尤其在关键决策时，仍然需要人的参与。这里举两个例子。
在流程 A 中，如果 Agent 检测到登录失效，就需要暂停当前流程，通知用户，并说明阻塞原因。用户收到通知后，可以接入云电脑，完成相应处理。随后，Agent 重新读取会话状态，继续执行任务。
在流程 B 中，遇到需要人查看现场状态的情况，Agent 同样可以通知用户，由用户介入处理，再继续完成任务。
从让用户理解当前状态，到在相应节点保存现场，再到人工接管后恢复任务，这些都需要电脑层面的能力支持。并不是大模型或 Agent 本身具备一定能力，就足够了。
接下来介绍三个过去一年中比较典型的客户案例。
第一个是利星行。利星行是全球最大的梅赛德斯-奔驰经销商，在全球有很多经销门店。
他们通过云电脑统一管理员工使用的 Agent、办公系统，以及 Token。员工无论使用手机还是平板，都可以方便地接入云电脑，使用其中内置的系统级 Agent 完成工作。
比如，门店员工需要制作视频时，可以把任务提交给 Agent。对于一些重复性工作，也可以通过云电脑与 Agent 协同，录制相关技能，进而自动生成报表。
无影在其中主要提供基础设施支持，包括数据中心、安全管理和 Agent 托管等。
第二个案例是好未来的“九章龙虾”。他们把 Agent 运行在云电脑中，也就是把“龙虾”养在云电脑里。
不到两周，他们就在企业级云电脑中部署了 1,000 只“龙虾”，为老师提供教案、课件制作，以及作业批改等能力。目前，这一规模还在扩大。
每位员工也可以使用无影云电脑进行调试。云电脑提供的数据恢复等能力，也能帮助他们做好数据安全方面的工作。
第三个案例是云道智能。他们使用了无影的 Agentic Computer 能力。
他们的产品是一款类似钉钉 A1 的硬件设备，接入了云电脑能力。设备中的 Agent 可以调用云上的云电脑，完成复杂工作。
这款设备在很短时间内就卖出了两万多台。最终用户可以使用云上的 Agent，以及 Agent 所在的云电脑，完成复杂任务。
数字世界也可以分成两部分：互联网和移动互联网。在互联网中，我们主要通过电脑和浏览器进行交互；另一部分，则是移动互联网。
最近 Personal Agent 很受关注，在座可能已经有人测试过千问 Go。它集成了一台无影云电脑，也集成了一台无影云手机，可以借助这两个工具，与互联网和移动互联网进行交互。
这里的视频是一个 Demo：用户提出任务后，Agent 可以自动操作云手机。视频展示的是机票退费的例子，由 Agent 自动完成相应操作流程。
过去五六年，云电脑最初主要是供人使用的办公工具，解决了身份和权限管理、资源管理，以及团队协作等问题。
进入 Agent 时代，Agent 所需要的“工位”，同样离不开这些能力：权限如何管理、身份如何管理、资源如何方便地创建。这些需求，与云电脑已经提供的能力非常匹配。
随着 Agent 不断演进，无影的云电脑、云浏览器和云手机等产品，也正在被越来越多的 Agent 使用。比如，内部的千问 Go、Qoder 等，都已经集成了这些能力。
如果你正在做 Agent 创业，或者你的企业正在使用 Agent，无影云电脑、云手机可以帮助拓展 Agent 能完成的任务范围。
原来的工作类 Agent 可能主要调用浏览器。现在，通过 MCP 或 CLI 接入云电脑后，Agent 就可以使用更加复杂的软件和业务系统。
最后，也邀请大家到阿里云官网了解无影云电脑、云手机和云浏览器。这些产品都已经针对 Agent 时代进行了深入改造，可以方便地接入 Agent，拓展它们的能力。
无论你使用的是 Claude Code、Codex，还是 Qoder，都可以通过标准的 MCP 方式接入相关能力，让 Agent 与大模型更好地服务业务。
我今天就分享到这里，谢谢大家。
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。
• At the main forum of 1024 Moot Ai Community Days, Alibaba Cloud ' s Vice-President for Research and Development, Head of the Department of Terminal Intelligence Computers, Zhang Tsi, shared the new role of cloud computers in the Age of Agent under the theme " The evolution of cloud computers: from human desktops to AI agent " : providing AI agent with a working environment that is sustainable, permission can control and work with people.
Zhang Tsai joined Alibaba Cloud in 2014 and has long been responsible for flexible computing services ECS, and then moved to business exploration for integration of terminal equipment with AI. From supporting cloud operations to providing an operating software environment for Agent, his concerns have also been extended: how can it work smoothly when AI agent begins to enter the enterprise workflow?
In the past, the cloud computer was a human teleworking tool; now Agent also needs such a “workspace”. It has to open software, access business systems, process tasks, encounter steps that cannot be continued and be taken over.
Give Agent a “position”, first of all, a tool to use. Previously, Agent had been performing tasks mainly through browsers; with access to cloud computers, they could further use desktop software and enterprise applications to handle complex work across systems. This capability is particularly important for old systems that do not have open API and still rely on a graphical interface.
But having a computer doesn't mean that Agent can work in the business. It needs to be clear as to the identity in which it enters, what data it can access, what operations it can perform, and how the process is documented and audited. The role of cloud computers also extends from providing an operating environment to the integrated management of identity, permission and resources.
Another key question is: Agent's stuck. Can you get it?
Zhang Tao took the example of order processing: after Agent received a new order, it had to check the information, generate a report and submit it to the supervisor for approval. This process appears to be smooth and may actually be performed in situations where login failure, authentication codes or manual judgement is required. Any step that could be blocked could stop the mission.
At this point, cloud computers are required to maintain the current working state, notify users and explain the reasons for the jamming and allow access to the processing. After the problem is solved, Agent goes on to execute. Such collaboration requires both an understanding of the steps to be taken in the task and a systematic presence to avoid taking over from scratch.
These capacities have already been utilized for specific operations. According to Zhang Jingtao, within less than two weeks, a project for the future has deployed about 1,000 Agents through enterprise cloud computers to support teachers with teaching, curriculum production and job modification. Staff can also access cloud computers for calibration and use capabilities such as data recovery to secure data.
In sum, according to Zhang Jingtao, the cloud computer has accumulated many years of identity permission management, resource management and collaborative capabilities, and can happen to be the “position” of Agent: it provides Agent with an environment for running software, accessing business systems, and managing identity and permission. When Agent encounters problems, a person can access the same cloud computer for processing and then continue the mission by Agent.
The following is a complete summary of the presentation, which was revised and edited without changing the original intent of the workshop.
Agent also has a "position": the cloud computer allows it to take over the business and controls permission.
Ladies and gentlemen, good holidays! Welcome to Cloud Valley, and thank Kevin for his invitation. I've had a lot of communication with Kevin over the past few years. It was also a pleasure to meet many old friends today, as well as Bigen.
The first sentence after Begen met was, "Your cloud computers are now the infrastructure of Agent, and should be sold well?" I said, "That's true."
Let's introduce ourselves to our new friends here. My curriculum vitae is relatively simple: I worked in Intel nine years before graduation, mainly in research and development related to core cloud computing techniques. Joined Alibaba Cloud in 2014, it's been 12 years now. Ten years ago, I was primarily responsible for flexible computing, the ECS. At the beginning of 2024, I started in-house entrepreneurship, exploring new business with terminal equipment and AI.
Over the past few years, we've made a shadowless cloud computer. We also asked Book, the first AI agent computer in the industry, at the Cloud Habitat Congress.
The topics discussed today are closely related to this: The Shadowless Cloud computer ushered in a new era, and it was becoming the “position” of AI agent.
First, a review of the changes over the past three years. At the beginning of 2023, this round of AI waves, represented by LLM, broke out in approximately three phases.
The first stage, represented by ChatGPT like Chatbot. It is a shock to everyone that artificial intelligence is indeed powerful and has been able to easily handle natural languages and communicate with human beings.
Phase two, represented by Agent applications, such as Coding Agent and Work Class Agent. They show us that AI can use tools to replace human beings in very complex tasks.
The third stage is the recent focus of personal Agent, an individual AI agent. Various manufacturers are involved in publish-related products, and Ali also asked publish about Go (the Personal Agent, launched in 2026).
This means that Agent is entering a very important scene: personal life and work.
One of the things that has been repeatedly stressed by friends who are concerned about the personal Agent is that each Agent is equipped with a cloud computer, so that Agent can perform complex tasks to the cloud computer. Previously, it might have been primarily a browser; with cloud computers, more complex work could have been done.
The rationale behind this is simple. The world that we touch can be divided into physical and digital worlds. The physical world is the real world that we can reach, and the digital world includes the Internet and mobile Internet, which is built by human beings.
Agent, if he really wants to help human beings, who represent human beings, to do more complex work, needs to use the tools that humans use to deal with the digital world.
So, as Bigen just said, the cloud computer became Agent's infrastructure. We usually joke that this is a good "position" for Agent.
And look at the evolution of the cloud computer.
The first generation cloud computer was more present in the call centre, usually called VDI. Many companies deploy VDI or cloud computers in data centres.
In the second phase, cloud computers can already complete more complex tasks. In the case of Shadowless Clouds, for example, we have served hundreds of thousands of businesses, with millions of users working on clouds every day.
This is of great value to both businesses and individuals. For enterprises, cloud computers help better manage IT resources and access internal systems. When business systems and office systems are all clouded, efficiency gains will be significant.
For employees, access to the company ' s systems is easy and safe, whether by mobile phone, tablet or web page.
With business developments and load changes, we have also launched GPU workstations over the past few years. Three-dimensional design, simulation, and training and reasoning for AI models can be done through such workstations.
As these capabilities build up, and as they enter the AGI era, the cloud computer opens up new opportunities for development. It can accept the task of Agent to get Agent to access the complex software in the computer.
For example, LLM can perform video production, 3D production, etc. through the Computer Use operating software.
In this scenario, cloud computers become more important. What capacity is needed to meet Agent's needs? These are also the capabilities that we have accumulated over the years in the field of cloud computers.
First, the cloud computer is not just a cloud computer, but also provides identity and security management, including authentication, auditing, and access to various systems permission management.
Secondly, there is interaction and system capacity. Through a self-study agreement, users can fully use the cloud operating system, install software and have clear, fluid interaction.
Thirdly, resources and life cycle management. Enterprise procuring entities usually require a certain period of time for computers; but in the clouds, only a few orders are required to create thousands of computers, tens of thousands of computers and complete life cycle management.
How do these capabilities respond to the computer requirements of the Age of Agent? We made a further comb.
First, there is “workspace”. Cloud computers can provide Agent with a secure, secure and stable cloud environment of 7x24 hours and readily accept Agent's mission requirements.
Secondly, control. Our ability to identify, authenticate, and so on, enables Agent to have his own digital identity when accessing the enterprise system, namely, Agent ID, and to perform complex tasks within the enterprise under the corresponding permission.
Thirdly, there is synergy. In carrying out its mandate, Agent often requires the involvement of human-in-the-loop. It still needed to hand over control to the person in case of problems that could not be resolved independently, such as certification codes, or when it needed to log in to a system.
Cloud computers already have the capacity to do so. After Agent tells you, you can access it with different devices and work with Agent.
Why does Agent need a sustainable environment? We use an example.
Many manufacturing firms have ERP systems, and the process is relatively simple for people. Assume that Agent will be given a time job: When new orders are made on a daily basis, the order and related system information are automatically verified, a report is generated and forwarded to the supervisor for approval.
This process may encounter a variety of anomalies, resulting in Agent not being able to perform its tasks automatically and requiring manual intervention. In the case of problems, how Agent interacts effectively with people is also an issue that needs to be addressed.
These capabilities, cloud computers, have accumulated over the years. Some systems can already be operated through CLI, others can be called directly to API, but others, especially older ones, still need to be operated through a graphical interface.
In this complex cross-system, cross-application task, synergy between people and Agent is important. Without the support of a good cloud computer, it would be difficult for Agent to do such a job.
In addition, Agent faces many challenges when entering the enterprise ' s production system or environment.
For example, having Agent represent you every week to produce a weekly newspaper, which may also be subject to anomalies. How complex missions are handled and how permission is managed between systems is not a computer that can be solved.
Enterprises also need to harmonize their management and regulation so that Agent has its own ID and the corresponding permission to connect to different systems and perform complex tasks.
Agent also needs to interact with people, especially in key decision-making processes. Here are two examples.
In Process A, if Agent detects login lapses, it is necessary to suspend the current process, notify the user and explain the reason for the blocking. Upon notification, the user can access the cloud computer and complete the corresponding processing. Subsequently, Agent rereaded the session and continued its mission.
In process B, in cases where a person is required to view the situation on the ground, Agent can also notify the user, process it with the user and continue the assignment.
Computer-level capacity support will be required from users to understand the current state, to maintain the site at the respective nodes and to resume the mission after manual takeover. It's not that LLM or Agent has some power, that's all.
This is followed by three more typical client cases over the past year.
The first is the Star of Leigh. It is the largest Mercedes-Benz dealer in the world, with many distribution stores worldwide.
They centrally manage Agent, the office system and Token for staff through cloud computers. Both mobile phones and tablets provide easy access to cloud computers, using the system-level Agent embedded therein.
For example, when a door shop employee needs to produce a video, he or she can submit the task to Agent. For some repetitive work, it is also possible to record relevant skills through cloud computers, in collaboration with Agent, and thus automatically generate reports.
This includes infrastructure support, including data centres, security management and Agent hosting.
The second case is a good future, the Nine Zone Lobster. They put Agent in the cloud computer, which is the lobster in the cloud computer.
In less than two weeks, they deployed 1,000 lobsters in enterprise-level cloud computers, providing teachers with the skills to teach, to produce and to adapt their work. At present, this scale is expanding.
Each employee can also debug it using a cloudless computer. Capacity such as data recovery provided by cloud computers can also help them to do their job in data security.
The third case is cloud intelligence. They used the shadowless Agenic Computer capability.
Their product is a piece of hardware similar to nail A1, which connects to cloud computer capability. Agent in the device can call the cloud computer and do complex work.
More than 20,000 units were sold in a short time. The end-user can perform complex tasks using the cloud Agent and the cloud computer where Agent is located.
The digital world can also be divided into two parts: Internet and mobile Internet. In the Internet, we interact mainly through computers and browsers; in the other part, mobile Internet.
Recently, Personal Agent has received attention and may have been tested for Go. It combines a shadowless cloud computer, as well as a shadowless cloud cell phone, which allows interaction with the Internet and mobile Internet.
The video here is a Demo: After the user's job, Agent can automatically operate the cloud phone. The video shows examples of airline ticket refunds, which are automatically completed by Agent.
Over the past five to six years, cloud computers have initially been primarily a human-use office tool, addressing issues of identity and permission management, resource management and teamwork.
To enter the age of Agent, the “positions” that Agent needs are equally necessary: how permission manages, how identity is managed, how resources are easily created. These requirements are well aligned with the capabilities already provided by cloud computers.
As Agent continues to evolve, products such as dark cloud computers, cloud browsers and cloud cellular phones are being used increasingly by Agent. For example, the insider, Go, Qoder, and so on, have been able to build these capabilities.
If you're doing the Agent start-up, or your business is using Agent, Shadowless Clouds, Cloud Cell phones can help expand the range of tasks that Agent can do.
The former working class Agent may call primarily the browser. Now, with access to cloud computers through MCP or CLI, Agent can use more sophisticated software and business systems.
Finally, you are invited to visit the Alibaba Cloud official network to learn about shadowless cloud computers, cloud phones and cloud browsers. These products have been further adapted to the Age of Agent, allowing easy access to Agent and expanding their capabilities.
Whether you use Claude Code, Codex or Qoder, you can access the capabilities in a standard MCP way to make Agent and LLM better service the business.
I'll share it with you today. Thank you.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.