Betting on Agents: The Midfield Battle of AI among Tech Giants
While you're still questioning the practical value of large AI models, AI agents have swept in, triggering a new technological wave globally. Leading the charge is Manus, a platform touted as the world's first general AI agent.
Developed by a Chinese team, Manus has been hailed by some as the best AI agent currently available, a black - tech marvel like something from outer space. Its birth is regarded as the "DeepSeek moment" for AI agents.
Manus utilizes a multi - agent collaboration architecture to handle tasks automatically. For instance, if you need to process ten resumes, you can upload them to Manus. It will automatically read the resume content, conduct data analysis, and generate a report. While it's helping you work, you can put down your computer, grab a cup of coffee, and come back to check the report.
In daily life, if you plan to buy a house in a certain city, previously you had to painstakingly search for housing information and compare prices. With Manus, it can break down the "house purchase" into smaller tasks, write Python scripts to calculate the budget, and even generate a detailed report like that from a real - estate agency.
In comparison with AI agents, large generative AI models that adopt the "question - answer" mode and highly rely on users' prompts seem less flexible.
Is Manus really as great as the outside world claims? Are AI agents a genuine need or just a gimmick? How are domestic and international tech giants reacting to this new technological wave?
1. Manus: Different Fates in Domestic and Overseas Markets
A few months ago, Manus became extremely popular in China, and its invitation codes were once resold for tens of thousands of yuan.
On May 12th, Manus opened free registration globally. Everyone can use it without an invitation code and get 1000 credits at once per day to execute tasks for free.
To sustain itself, Manus offers users more access rights, exclusive features, and priority services through three tiers of paid subscription plans, attempting to recover costs through membership fees.
Meanwhile, Manus is constantly evolving. On June 4th, it launched the "AI text - to - video" function. Users only need to input a few key prompts, and the system will automatically plan each scene and create visual effects, trying to compete with video - generation platforms like Sora.
Although Manus is regarded as a technological myth by its domestic and international supporters, the domestic market doesn't recognize its efforts.
Manus has been involved in marketing controversies. Some believe that the previously scarce test invitation codes were part of a carefully - planned hunger - marketing strategy.
What's most unacceptable to the domestic market is that Manus is a "wrapper" product. The domestic AI market always values self - developed large models. Manus doesn't have its own large model; it uses Anthropic's Claude overseas and Alibaba's Tongyi Qianwen in China. Its functions are pieced together from multiple applications. In most people's eyes, this means giving up its core competitiveness and having no moat.
However, for Silicon Valley overseas, Manus' "wrapper" nature is actually its advantage. Since Manus solves the last - mile problem of AI, even without researching the underlying models, innovation at the application level is very important. Overseas investors consider Manus an important technological advancement.
Manus' progress is based on a revolutionary technological upgrade of intelligent agents through the CodeAct framework, endowing it with three core capabilities. Firstly, it can directly execute Python code to complete highly complex tasks. Secondly, it can dynamically adjust action strategies in real - time according to task requirements. Thirdly, it realizes the self - regulation and optimization functions of agents. Essentially, this is a technological innovation that unlocks new AI application directions. Manus has not only quickly secured financing and thrived overseas but also gained numerous supporters.
Manus' founding team seems to follow the "overseas first" script. The founder sits in front of the camera, fluently introducing in English what general tasks their AI agent can perform and showing specific examples on the screen.
The International Data Corporation (IDC) predicts that the domestic artificial - intelligence market will reach $26.44 billion in 2026, a growth of about 17.9% compared to 2021. The global artificial - intelligence market is expected to reach $301.43 billion in 2026, with the domestic market accounting for about one - eleventh of the global total. It seems an irrational choice for the Manus team to focus on the overseas market while ignoring the huge domestic market.
However, the main reason for this might be that as a startup, Manus doesn't want to compete with domestic AI giants.
Without a self - developed large model, Manus has high costs. Its product pricing is restricted by large - model manufacturers, and the price can't be lowered, making it unable to compete with the free AI agents of domestic giants.
In China, domestic giants started aggressive layout as soon as Manus became well - known.
2. Coze Space Leads the New Wave of AI Office Work
On April 18th, ByteDance's AI agent product, Coze Space, caused a sensation as soon as its internal testing began.
A large number of enthusiastic users overwhelmed the server instantly, and invitation codes were extremely hard to get, perfectly replicating Manus' popularity. The success of this product proves the market's strong demand for AI agents that can solve specific work - scenario problems.
Coze Space can write reports, search for information, format PPTs, and even build websites. When you assign a complex task to it, it can break it down into smaller steps and execute them on its own. The process is simple and fast, meeting the needs of office workers. This is also Coze Space's product positioning: "Start your work with the agent."
Coze Space has two modes: exploration mode and planning mode. In exploration mode, it can automatically execute tasks according to user needs. In planning mode, after the user puts forward a demand, it will first provide a task - processing plan and then start acting. This way, the interaction between the AI agent and the user can better solve problems.
According to test results, Coze Space is like an intern at work. Users can assign simple tasks to it, and it can collect information and deliver results just like a human.
If you ask Coze Space to create a plant - science popularization game, it can give you a web - based game that includes weather, plant information, and card collections. You can get popular - science knowledge by clicking on the plant information. Additionally, it will provide an interactive web file, allowing users to observe the whole process and jump to any step to view detailed content.
Coze Space doesn't just aim to be an intern for regular users; it wants to be an expert in niche fields and meet the needs of in - depth users.
This provides more operating space, diverse user - selectable modes, and improves work productivity. Coze Space officially demonstrated its capabilities in market - research analysis, customized stock morning reports, and interactive teaching, covering a total of more than a dozen work and life scenarios.
However, Coze Space is not perfect and is still an immature product. After all, users may encounter situations where the tasks they demand are too complex to be achieved, and it can't fully meet users' personalized needs.
In the context of the continuous evolution of AI technology, 2025 is regarded as the first year of AI agents. The launch of Coze Space is the vanguard for large companies to enter the AI - agent competition.
3. Baidu Miaoda's Breakthrough
Baidu is trying to find a way out in the fierce competition in the AI - agent field. On March 24th, Baidu held a grand press conference in Beijing, announcing that its self - developed generative AI application platform, Miaoda, was officially open for commercial use.
The question is, there are more than a dozen companies developing programming tools in the AI circle. What makes Miaoda special?
The answer lies in three words: "no - code".
Other AI programming tools require users to have some coding foundation for subsequent modifications. Miaoda, however, doesn't require users to understand code. As long as users can express their ideas, they can have the same programming ability as programmers and develop the required applications smoothly.
Robin Li said, "As long as you have an idea, you can make it come true. We're entering an unprecedented era where you can make money just with ideas."
This statement is based on Miaoda's three characteristics: no - code programming, multi - agent collaboration, and multi - tool invocation.
Firstly, no - code programming, based on the development ability of the Wenxin large model, makes programming accessible to everyone without a threshold, subverting the traditional programming model. The multi - agent collaboration architecture automatically breaks down complex tasks and coordinates the scheduling of different agents. Its advantage is to improve the development efficiency and quality of complex AI applications. The multi - function invocation ability integrates tool services within Baidu's ecosystem and can call professional tools such as web search, maps, and document analysis. Users of Miaoda only need to state their needs, and the system will provide the best results.
In simple terms, describe your needs in Chinese, and Miaoda will provide a programming application.
If Luobo Kuaipao wants to organize a press conference but doesn't know how to create an online registration system, it can use Miaoda. After describing the specific needs and uploading a document with the conference time, location, and theme, Miaoda will give you a perfect press - conference registration system application.
However, many users found that after using Miaoda, it's not the all - powerful perfect product as advertised.
Miaoda performs well in simple game applications and website generation. But in complex applications like intelligent quiz systems, due to its limited understanding ability, it requires a large number of prompts for debugging, and the operation is cumbersome.
Especially in the modification of complex applications, it can only be modified through dialogue and the source code can't be downloaded. This is a significant drawback for users. If there are problems with the generated product, Miaoda's problem - solving ability is questionable, and it also makes it difficult for users to interact with the server.
Overall, Miaoda's functions are relatively powerful. For users whose main occupation is programming, they can use Miaoda as a reference when there is no UI, but it can't be used as a production tool. The results it produces are just at a passing level, and it's mediocre overall, having some merits but not outstanding.
4. Quark and DingTalk: Alibaba's Dual - Front Battle
While its peers are still discussing the form of agents, Alibaba has already let agents create real value in business practice.
First, Quark has been upgraded from a simple browser to the "new Quark", a multi - faceted flagship application centered around AI. Its flagship Super Box function breaks the previous simple search mode. As long as users describe their needs in this "box", they can get reports, guides, and plans completed by AI through invoking various tools. It's an AI agent that can deliver results according to user needs.
Quark had a poster during its promotion, saying "Farewell to Search". This is Quark's core competitiveness. The purpose of using the Super Box is no longer just a simple search tool; it's an AI product that can help users solve problems and can link all of Quark's internal tool capabilities to serve users.
As an application labeled "All in One", in Quark's Super Box, users can not only input text but also take photos or input voice, simplifying the user - operation mode.
Backed by the Alibaba ecosystem, Quark can access Taobao shopping data, Gaode location information, Alipay transaction records, etc., providing users with highly personalized services.
Then there's DingTalk. If Quark is a C - end product for ordinary consumers, DingTalk is a standard enterprise platform.
DingTalk launched the AI Agent Store last year. This platform initially launched more than 200 vertical - field AI assistants, covering applications in multiple scenarios such as enterprise services, industry applications, and life entertainment, constructing the most complete domestic B - end AI application matrix.
DingTalk has accumulated rich scenarios and industry data from various industries and has clear customer needs, having obvious advantages in ToB - end AI agents. This strengthens the agent's long - term and short - term memory capabilities, significantly improves the continuous tracking and execution ability of complex tasks, and shows strong cross - platform task - collaboration ability.
In the context of AI agents becoming the core track of global technological competition, Alibaba, leveraging its comprehensive advantages of "cloud + terminal + ecosystem", has constructed an AI - agent strategic layout covering both the consumer end and the enterprise - service end.
Alibaba is upgrading Quark to a "super personal agent" on the C - end and transforming DingTalk into an "enterprise - level AI agent platform" on the B - end.
Alibaba's core in the C - end market is Quark. This product, originally positioned as an intelligent search engine, has gradually been upgraded to a personal intelligent assistant integrating AI capabilities. In the B - end market, Alibaba uses DingTalk as the core carrier to construct the most complete domestic enterprise AI - agent matrix.
However, in the C - end market, the popularization of agents still faces significant challenges. The emergence of the large - language model ChatGPT earlier had an impact on the AI - question - answer field, making ordinary people more accustomed to the concept of "AI". Most consumer - oriented agents are still in the "trial phase", and users can't yet understand the value of agents.
In contrast, the development of AI agents in the B - end market is smoother. Since enterprises usually have structured data like internal knowledge bases, agents can execute tasks more accurately, and the delivery results are easier to quantify. Enterprises are more willing to pay for measurable efficiency improvements.
The development of AI agents is reshaping the competition logic of the domestic Internet industry. For domestic leading enterprises, this is both a strategic opportunity to overtake on a curve and a transformation test in the AI - intelligent era.
When Meituan Waimai's AI can handle customer complaints independently and Douyin's creation AI can generate millions of short - videos in batches, Internet companies need to rethink their essence. Maybe there will be no pure "Internet enterprises" in the future, and all surviving companies will be "AI - agent operators".
This article is from the WeChat official account "Jia Bin Shang Xue" . Author: Jia Bin Shang Xue. Republished by 36Kr with permission.