Byte, Alibaba, Baidu and other major manufacturers have entered the intelligent body track and launched products to participate in the traffic entrance battle, but they have not yet broken through the boundaries of capabilities. Although Byte’s “Button Space”, Alibaba’s “Flow”, Baidu’s “Heartbeat” and other products have their own focuses, they all face the dual anxiety of technology replication and traffic competition, and the future battle of agents is still full of uncertainties.
When Xiao Hong, the founder of Manus, bluntly said that “there are no secrets in products”, large manufacturers are caught in the double anxiety of technology replication and traffic competition. According to Lightcone Intelligence, many major manufacturers have set up internal product teams to benchmark Manus. It is reported that Byte, which is good at internal horse racing, has at least 5 different teams developing agent products internally.
A month after the release of Manus, major manufacturers have handed over the answer sheets of agents: ByteDance’s “Button Space” has swept the workplace with fission invitation codes, Alibaba’s “Flow” has shaped the depth of research with ultra-long time-consuming tasks, and Baidu’s “Heartbeat” has raided the mobile terminal with medical and legal vertical scenarios.
But behind this seemingly prosperous wave of generic agent releases, there is an embarrassing reality – all products have not yet broken through the boundaries of existing agents.
From opening the Agent platform last year to handing over 60 points this year. What is certain is that the Agent has crossed the 1.0 stage of the General Agent through planning and tool capabilities and entered the 2.0 stage of the Autonomous Agent.
At present, the Agent released by Zudui has been able to initially take on the task of “intern” to solve those time-consuming and low-difficulty tasks in human daily life:For example, according to the needs of users, a part of the data is collected, and some views can be filtered through a large amount of information analysis. Another example is to help users complete basic operations, screen positions, send resumes, send emails to targets, etc.
This battle for traffic entrances around Agents is just a footnote to the year of the outbreak of Agents in 2025. In addition to large model companies and application companies, more players are participating:
Just today, Lenovo also released the “Tianxi” agent for individual users and the “Lexiang” agent for enterprises. Among them, the personal agent “Tianxi” takes autonomous operation and memory interaction as its selling points, and is equipped in its 4 products. In addition to autonomously completing orders, strategies and other operations, it can also give matching budgets and preferences based on users’ preferences and habits.
How can product managers do a good job in B-end digitalization?
All walks of life have taken advantage of the ride-hail of digital transformation and achieved the rapid development of the industry. Since B-end products are products that provide services for enterprises, how should enterprises ride the digital ride?
View details >
Under the attack of the big factories exhausting their resources, what stage has the battle of agents reached now?
01 Batch replication of Manus, who can seize the traffic entrance?
In April, GM Agent became a key player in the press conferences and product updates of major manufacturers.
First, Byte released the general agent product “Button Space” on April 18, launched a free test, and seized the market first by fission method of 5 invitation codes per person; On April 22, Alibaba’s AI assistant “Flow” launched advanced research functions, focusing on DeepResearch-style report research similar to OpenAI’s agent; and then Baidu officially announced the agent “heartbeat” at the developer conference on April 25, focusing on medical, legal, picture books and other task experiences adapted to life scenarios.
After the impact of DeepSeek, how to drain traffic with the help of mature products has also become a problem that needs to be considered in the next launch of the agent.
At present, most large factories are in the mode of self-development by internal groups, such as Byte and Baidu. But Alibaba had already negotiated cooperation with Manus as early as when Manus ignited market enthusiasm. At present, in addition to the release of its self-developed product flow, its large model team Tongyi Qianwen has also been finalized as early as March to become the exclusive model supporter of the Chinese version of Manus, and will first get a piece of the model cooperation.
However, for Baidu, Byte and Alibaba, the three major manufacturers that emphasize AI self-development and do everything from large model research and development to application products, it is also very necessary to develop agents with their own strength.
The Agents who released together have different product ideas.
The “buckle space” of the byte is positioned as “Interns who are proficient in various skills”The various cases given are more like work assistants, such as generating industry reports, user research and analysis, etc.
After the intelligent test of the light cone, I feel that the buckle space is a more comprehensive passing product, which not only connects to AutoNavi and other MCPs, but also improves the ability to use with the help of tools, and also achieves pictures and texts in the report output, and the multi-modal ability is outstanding. However, there is room for further optimization in the depth of the output report.
In contrast, the advanced research function launched by Alibaba’s AI assistant “Flow” is closer to the application scenarios of in-depth research. During multiple tests, the significant feature of flow is that it consumes more token resources and takes a long time. For example, in the “agent research” task, the processing time of flow exceeds 1 hour, and the number of web browsing also significantly exceeds that of the other two products.
A lot of resources and a way to sacrifice efficiency in exchange for more in-depth generated content,And this is also the reason why flow can only pass the review system application test, large-scale opening, for computing power consumption and cost, are currently difficult to balance. For example, in the agent report, most of the products it chose were large model AI assistants, and the information of Manus’ financing of 75 million yuan was mistakenly placed in the analysis of the buckle space.
Compared with Byte and Alibaba’s general agent choice to launch on the PC side, Baidu’s heartbeat chose the first launch on the mobile app side, and then it will be launched on the PC side.
Different intelligent hardware ends determine the difference in audiences faced by the two: most of the computer side faces users with work and study needs, focusing on the fields of reporting research and content analysis; Mobile phone users are more eager to experience the functions of AI in real life scenarios.
Combined with Baidu’s advantages in medical, legal and other industries, coupled with Baidu’s past exploration results in AI virtual social networking, Xinxiang’s final online form has become the appearance of the main vertical scenario.
At present, the main interface of the Xinxiang App recommends experience recommendations including AI blind dates, travel planning, medical/legal consultation and other scenarios. After actual testing, the accuracy of heartbeat in medical, legal and other issues is high. For example, when answering the question of drunk driving hitting and fleeing, Xinxiang called multiple agents and finally gave a sentence of 7 years, which is also consistent with the opinion of lawyers in reality.
However, in other businesses that are separated from the accumulation of vertical knowledge, the accuracy of heartbeat needs to continue to be optimized. For example, in the task of requesting a tourist location suggestion, Xinxiang has clearly judged that the location does not belong to the scope of Chaoyang District, but still recommends attractions that do not meet the conditions set by users.
Based on the evaluation results of the three general agents, most of the AI agents in the echelon of large factories can only be used to complete some basic level work, and their capabilities have not yet reached the amazing effect of Manus.
But in the short term, the answers handed over by major manufacturers have made market users gradually excited and curious about the concept of agents.
Whether it is Byte or 360’s Agent, there have been server crashes in the process of large-scale open testing recently, and it can be said that even large manufacturers with sufficient computing power have far exceeded the expectations of the release.
In the track of agents, the participants are not only Internet giants and startups, but also companies with smartphones and computers with hardware advantages are also eyeing the prospects of agents.
Today, Lenovo released the “Tianxi” and “Lexiang” enterprise super agents for individuals at the press conference, and for individuals, its agents have been able to complete various autonomous operations such as providing travel suggestions, making itineraries and placing orders.
In the face of the fat of “agents”, all companies are eyeing it.
02 Manus has not been surpassed yet, but how long can the technical dividends be eaten?
In 2025, known as the “Year of Agents”, large factories and startups will hand over answers within their capabilities.
However, the products currently launched by the three major manufacturers reflect a cruel reality: even if the big manufacturers have ecology and computing power, replicating Manus is not something that can be done in a short period of time.
At the beginning of the release of Manus, its founder Xiao Hong had already given his own opinion: Manus has no secrets.
“From the perspective of product managers, if you want to use it, you must use the best large model, and how much commercial value can be generated by using the best things, product managers are concerned about this.” Xiao Hong said.
Manus’ core capabilities are based on model capability overflow, but at the earliest moment of establishing market awareness, Manus has achieved the ultimate in model capability call and product ideas:
First, the agent needs to handle diverse tasks such as multimodal understanding, complex reasoning, and code generation at the same time, which puts forward extremely high requirements for the comprehensive calling ability of the underlying model.Most of the general-purpose agents that can be seen on the market now use not just one large model, but multiple models to call according to different needs, such as calling a large model with stronger multi-modal capabilities if you need to understand the content of the picture.
Taking Byte as an example, according to LatePost, when the buckle team developed the buckle space, they considered using DeepSeek-R1 first, but after testing, they found that its ability to call the tool was insufficient. Finally, consider based on capability performance and cost reasons. It uses a variety of models, mainly the Doubao 1.5 Pro.
On the basis of better model capabilities, how to transform technology into user experience is also a challenge.
However, on the agents handed over by each company, different ideas have been shown. For example, in terms of search experience, Baidu’s Agent product adopts the “multi-set keyword + search engine” strategy to try to search with multiple sets of keywords, while Zhipu allows its own AutoGLM to contemplate searching on different platforms such as Xiaohongshu and Zhihu, breaking the boundary of not applying data solidification.
In the short term, the current technology landscape presents an interesting phenomenon: startups represented by Manus and GensPark are still leading in key indicators. For example, GensPark has outperformed Manus in the GAIA benchmark.
In contrast, the Agent products launched by large manufacturers at this stage are more “60-point solutions”, but with the same goal as Manus, large manufacturers need to take the lead in seizing part of the market in the relatively blank stage of general agents.
The difference in technical routes directly affects the choice of commercialization strategy, and the two types of players are moving towards completely different paths:
At present, mainstream manufacturers are still taking the free strategy as the leading goal, trying to pry away some users, and Byte, Alibaba and Baidu’s products have all adopted the method of free supply.
Cost pressures force startups to start commercializing earlier. Currently, GensPark has accumulated 10,000 paying users and its ARR (annual recurring revenue) has reached $22 million. Manus offers two charging models in overseas markets: the basic version of $39 per month and the premium version of $199 per month, which are priced comparable to OpenAI members.
However, from the perspective of basic large models to application products, the window period for technical advantages is currently shortening, leaving little time for startups.
When the new general agent is raised to 70 points and 80 points, plus the free strategy, it is bound to have a further impact on startups.
The end of this competition is still in its infancy, and all the confrontations before the general agent ability jumps again are just the prelude to the outbreak of the 2025 Agent.
