阿里腾讯华为,国产AI扎堆做Codex,桌面Agent迎来大混战Nvidia vs. AMD: The Ultimate Graphics Card Showdown


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/FiO8ZrvHMqn8cIFmgJf5TQ
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

 

图片

上周和几个做C端AI产品的朋友吃饭,所有人都在吐槽同一个问题:旗下对话类AI产品的用户打开率已经跌到了8%以下,还不如一个普通的工具类APP留存高。

所有人都在问,下一个AI产品的爆发点到底在哪?答案其实已经摆在桌面上——没错,就是你现在正在用的电脑桌面。

拐点已至:AI从聊天框跑到了桌面

2026年3月OpenAI新一代Codex的爆火,彻底把AI产品的主线从对话交互,拽到了桌面环境任务执行的赛道上。

OpenAI官方测试数据显示,新一代Codex在Windows、macOS桌面环境下的标准化任务完成率达到89.2%,比2024年推出的GPT-4o桌面插件提升了47个百分点,单次任务的平均响应耗时压缩到了1.2秒,已经达到了商用落地的临界点。

简单来说,以前的AI是你问它“怎么做PPT”,它给你一堆文字教程,你得自己动手做。现在的桌面Agent是你说“给我做一份10页的Q4部门工作总结PPT,用公司统一模板,数据从上周的经营报表里提取”,30秒后一份改好格式、填好数据的PPT就躺在你的桌面上了。

图片

这个能力的突破,直接把AI的价值从“提供参考”变成了“替代执行”,整个C端AI产品的逻辑彻底变了。

国内竞速:抢滩入口之战

以前你打开电脑的第一步是开微信、开浏览器、开WPS,以后你打开电脑的第一步,是唤醒你的桌面AI助手,所有需求都通过它完成,中间所有的APP都变成了它调用的工具,根本不需要你手动打开。谁拿下了这个入口,谁就拿到了下一代互联网的第一张船票。

国内本轮桌面Agent的竞争,本质上是在争夺下一代用户数字生活的执行入口,替代的是过去十年里浏览器、办公软件、甚至操作系统本身的用户时长占位。

三条路径:当前桌面Agent的核心玩法

系统级嵌入:生态闭环的先手棋

华为鸿蒙NEXT、小米澎湃OS自带的AI助手,走的就是系统级嵌入的路线,直接把Agent能力做进操作系统底层。这类产品的核心优势是权限高,不需要跨应用申请授权,就能直接调用系统内所有软件的接口。

传统AI助手往往只是在现有系统上叠加的模块,而鸿蒙NEXT和澎湃OS直接将Agent嵌入系统内核。它们能够直接调用系统底层的数千项原生能力(如通信、日历、设备控制等),无需跨越多个APP的权限壁垒。

依托鸿蒙的分布式架构和澎湃OS的HyperConnect框架,Agent成为跨设备的“中控大脑”。用户可以在手机、平板、PC、车机、智能家居之间实现算力共享、文件流转和设备联动。用户只需表达模糊意图(如“我半小时后带朋友回家,准备一下”),Agent能自主决策、串联多个步骤(如调节灯光、播放音乐、关闭窗帘等)。

当然劣势也很明显:只能在自有系统生态内使用,鸿蒙的AI助手没法在Windows上跑,小米的也没法在macOS上用,用户规模天然受限于硬件销量。

图片

独立应用:跨平台的通用选手

阿里的千问、腾讯的workbuddy等这类独立客户端产品,走的是跨平台通用路线,Windows、macOS都能安装使用,不需要绑定特定硬件或者系统。

这类产品的优势是适配性强,用户下载就能用,功能覆盖也更全面,不管是办公做表、设计出图、还是查资料订机票,都能覆盖。截至2026年Q1,独立应用类桌面Agent的累计用户量已经突破2.3亿,占整个桌面Agent用户的78%,是目前市场渗透率最高的品类。

短板也很突出:因为没有系统级权限,很多操作需要用户手动授权,比如要读取浏览器下载的文件、要调取微信发消息,都得用户先点同意,一定程度上影响了使用体验。

现在各家都在疯狂和软件厂商谈合作,争取拿到更多的官方接口授权,本质上就是在补权限的短板。

图片

垂直场景深挖的差异化路线

本身有成熟工具产品的厂商,也是这波浪潮里跑在前面的一批。

比如WPS的AI助手,已经从最初的文档编辑功能,迭代成了跨应用的桌面Agent。用户给一句指令“把上周的销售Excel里的华东区数据摘出来,做成季度汇报PPT,同步发给部门群”,它就能自动打开对应Excel筛选数据,生成PPT图表,再跳转到企业微信完成发送,全程不需要手动操作。

还有做设计工具起家的厂商,推出的设计类桌面Agent,能对接PS、剪映、Figma等多类工具,自动完成批量修图、短视频片段拼接、封面图生成等任务。

图片

三大趋势:未来3年的行业演进方向

看完了现在的玩家和路径,我们可以清晰地推演出来这个行业未来3年的三个核心演进方向。

第一个趋势是交互入口去APP化。以前你要完成一个任务,得打开好几个APP来回切换,以后所有的任务都可以通过桌面Agent这一个入口完成,中间的APP全部变成后台调用的工具,根本不需要你手动打开。

未来3年,国内PC端用户的人均单日APP打开数量会从现在的11.7个下降到4个以内,80%的数字任务都会通过统一的AI入口完成。

第二个趋势是能力边界从“生成”转向“执行”。过去3年AI产品的核心能力是内容生成,写文案、画图、做视频,输出的都是半成品,还要用户自己二次加工。未来AI产品的核心能力是任务执行,你提需求,它直接给你交付最终结果,中间的所有操作都不用你管。

Gartner预测,到2028年,全球企业级AI产品里,任务执行类的占比会从2026年的17%上升到68%,AI Working会成为所有AI产品的标配能力。

第三个趋势是竞争核心从模型参数转向场景落地。前两年大家做AI产品,拼的都是模型参数规模、基准测试跑分,仿佛参数越高产品就越好。现在大家拼的是你家Agent能支持多少款软件的对接、能覆盖多少个场景的任务、能真正帮用户省多少时间。

未来桌面Agent的核心竞争力,从来不是你家模型有多聪明,而是你家模型能接入多少生态,能真正帮用户完成多少实打实的任务。

图片

对于所有科技行业从业者来说,这一轮桌面Agent的爆发,既是机会也是挑战。

做模型研发的,别再一门心思堆参数了,多花点精力做场景适配、做软件对接,比多堆几万亿参数有用得多。

做C端AI应用的,别再死守着对话APP不放了,赶紧往桌面执行的方向转,再晚就真的赶不上了。

做垂直行业工具的,赶紧把自己的场景做深做透,建立起自己的壁垒,不管最后哪个通用Agent赢了,你都有谈合作的筹码。

AI的下一个十年,早就不在聊天框里了,在每个人的桌面上,在每个能帮用户实打实解决问题的执行动作里。

图片声明:本文所用图片多数来自网络公开渠道,仅供分享交流使用,版权属原作者所有。如涉及侵权,请及时联系我们删除。

 

Source: https://mp.weixin.qq.com/s/FiO8ZrvHMqn8cIFmgJf5TQ If there is a copyright issue, please contact us at 13951431913#139.com or submit a comment on the right.

图片

Last week, I had dinner with several friends who are developing consumer-facing AI products. Everyone was complaining about the same problem: the open rate for their conversational AI products has dropped to less than 8%, which is even lower than the retention rate of ordinary utility apps.

Everyone is asking where the next explosion point for AI products lies. The answer is actually already on the table — it's your computer desktop.

Turning Point: AI Has Moved from the Chat Box to the Desktop

The breakthrough of the new generation of Codex by OpenAI in March 2026 has completely shifted the focus of AI products from conversational interaction to task execution in the desktop environment.

OpenAI's official test data shows that the new generation of Codex has a standardized task completion rate of 89.2% in the Windows and macOS desktop environments, a 47% improvement over the GPT-4o desktop plugin released in 2024. The average response time for a single task has been compressed to 1.2 seconds, which is the critical threshold for commercial deployment.

To put it simply, with the old AI, if you asked how to make a PPT, it would give you a bunch of text instructions, and you would have to do it yourself. Now, with the desktop agent, you can say, "Make me a 10-page Q4 department summary PPT using the company's standard template, with data extracted from last week's business report," and 30 seconds later, a finished PPT with the correct format and data will be on your desktop.

图片

This breakthrough has fundamentally changed the value proposition of AI from "providing suggestions" to "completing tasks," and has completely reversed the logic of consumer-facing AI products.

Domestic Competition: The Race for the Entry Point

In the past, the first step when opening a computer was to launch WeChat, a browser, or WPS. In the future, the first step will be to wake up your desktop AI assistant, through which all tasks will be completed. The individual apps will simply be tools called by the assistant, eliminating the need for manual opening. Whoever captures this entry point will win the first ticket to the next generation of the internet.

The current competition in China for desktop agents is essentially a fight to become the executive entry point for the next generation of user digital lives, replacing the screen time occupied by browsers, office software, and even operating systems over the past decade.

Three Paths: The Core Playstyles of Current Desktop Agents

System-Level Embedding: The First Move for an Ecosystem Lock-in

Huawei's HarmonyOS NEXT and Xiaomi's HyperOS have taken the route of system-level embedding, integrating the agent capabilities directly into the operating system kernel. The core advantage of this approach is high permission access. Since the agent is part of the OS, it can directly call the interfaces of all software in the system without needing to request authorization across applications.

Traditional AI assistants are often just modules added on top of the existing system, whereas HarmonyOS NEXT and HyperOS integrate agents directly into the system core. This allows them to call thousands of native system capabilities (such as communication, calendars, and device control) directly, bypassing the permission barriers of individual apps.

Relying on HarmonyOS's distributed architecture and HyperOS's HyperConnect framework, the agent serves as the "central brain" for multi-device connectivity. Users can share computing power, transfer files, and link devices between phones, tablets, PCs, cars, and smart homes. A simple intent, like "I'm bringing a friend home in half an hour, prepare everything," is enough for the agent to make decisions and coordinate multiple steps, such as adjusting the lights, playing music, and closing the curtains.

Of course, the disadvantage is obvious: the agent can only operate within its own ecosystem. For example, the HarmonyOS assistant cannot run on Windows, and the Xiaomi assistant cannot run on macOS. Therefore, the user base is naturally limited by hardware sales.

图片

Independent Applications: The Cross-Platform Generalists

Products like Alibaba's Qianwen and Tencent's Workbuddy are independent clients that follow a cross-platform strategy. They can be installed and used on both Windows and macOS, without requiring a specific hardware or system tie-in.

The advantage of these products is their strong adaptability; users can download them and start using them immediately. They also offer a comprehensive range of features, covering everything from office table creation and design drafting to information retrieval and ticket booking. As of the first quarter of 2026, independent desktop agents have accumulated over 230 million users, accounting for 78% of the total desktop agent user base, making them the most widely penetrated category.

However, they have a significant shortcoming: because they lack system-level permissions, many operations require manual authorization from the user. For example, to read files downloaded by the browser or send messages via WeChat, the user must click "agree" first, which degrades the user experience.

Currently, these companies are frantically negotiating with software vendors to obtain official interface authorizations, essentially trying to patch this gap in permissions.

图片

Vertical Scene Specialization: The Differentiated Route

Vendors who already have mature tool products are also leading the charge in this trend.

For example, WPS's AI assistant has evolved from simple document editing to a full-fledged desktop agent. A user can issue a command like "Extract the East China region data from last week's sales Excel, make a quarterly report PPT, and send it to the department group," and the agent will automatically open the Excel, filter the data, generate the PPT charts, and send the message on Enterprise WeChat, all without any manual interaction.

Similarly, companies specializing in design tools have created design-focused desktop agents that can connect with Photoshop, CapCut, Figma, and other tools to automatically complete batch photo editing, video clip splicing, and cover image generation.

图片

Three Major Trends: The Industry's Evolution Over the Next 3 Years

Based on the current players and their strategies, we can clearly extrapolate three core directions for the industry's evolution over the next three years.

First, the interaction entry point will become de-appified. In the past, completing a task required jumping between multiple apps. In the future, all tasks will be handled through the single desktop agent entry point, with the individual apps relegated to background utilities that are no longer opened manually.

Over the next three years, the average number of apps opened daily by PC users in China will decrease from 11.7 to fewer than 4, with 80% of digital tasks being completed via this unified AI interface.

Second, the capability boundary will shift from "generation" to "execution." For the past three years, AI products have focused on content generation (writing, drawing, video), producing semi-finished products that still require manual refinement. In the future, the core competency will be task execution. Users will simply state their requirements, and the AI will deliver the final result, handling all the intermediate steps.

Gartner predicts that by 2028, the proportion of enterprise-grade AI products focused on task execution will rise from 17% to 68%, with AI Working becoming the standard feature for all AI products.

Third, the competitive focus will shift from model parameters to scenario-specific landing. In the past, companies competed based on the scale of their model parameters and benchmark scores, assuming larger parameters meant a better product. Now, the competition is about how many software integrations an agent supports, how many specific scenarios it can handle, and how much actual time it saves the user.

The core competitiveness of desktop agents in the future lies not in the raw intelligence of the model, but in the breadth of their ecosystem integration and their ability to complete practical, real-world tasks.

图片

For all professionals in the tech industry, this wave of desktop agents represents both an opportunity and a challenge.

For those developing models, it's time to stop obsessing over parameter counts and focus more on scenario optimization and software integration, as these are far more impactful than just adding more layers to a neural network.

For those working on consumer-facing AI applications, the time has come to abandon the stand-alone chat app model and pivot toward desktop execution, or risk being left behind.

For those in vertical industries with specialized tools, the strategy should be to deepen the capabilities within your specific domain to build a solid moat. Whether a general agent wins the race, your specialized expertise will ensure you have a seat at the table for partnerships.

The next decade of AI is not happening in the chat box; it's unfolding on everyone's desktop, in the tangible actions that actually help users get things done.

Image Disclaimer:Most of the images used in this article come from public internet channels, and are for sharing and exchange purposes only. The copyright belongs to the original author. If there is any infringement, please contact us promptly for removal.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry