An der heißesten Weltkongress für künstliche Intelligenz (WAIC) der Geschichte vorbei: Sechs Trends bei Agenten

At this year's WAIC, the most frequently mentioned term is undoubtedly "Agent". More than 800 companies brought over 3,000 exhibits, doubling the number from previous years.

Not long ago, people were still debating whether Agent was a passing fad or a genuine trend. However, through this "hottest ever" World Artificial Intelligence Conference, the industry has realized that Agent is no longer a concept worthy of doubt. It is becoming an "essential question" for the industry, and an increasingly important one at that.

Trend 1: Agents are a must, and everything can be an Agent

Everything can be an Agent. Reporters witnessed this trend at the WAIC 2025 site. From industry giants to AI startups, from the "Six Rising Stars" to the "Big Dipper Seven", no booth could do without the keyword "Agent". The "neural network" of AI is spreading throughout the entire industry.

Stepping into the Siemens booth is like entering an AI factory. Engineers give the "add a new order" task to the Agent via voice, and the code seems to come to "life", constantly flickering on the screen. The robotic arms in the back also gain "autonomy" as the code is generated, picking up packages, transporting, and sorting them. Behind the efficient integration of this series of complex industrial processes, from order input, demand forecasting, equipment control to logistics scheduling, is a microcosm of the unified scheduling ability of Siemens' Industrial Copilot intelligent agent system.

"This is the first time the Industrial Copilot intelligent agent system has been unveiled in China. Its highlight is not only being a'super assistant' but also an 'industrial on - site conductor'," a Siemens staff member told a reporter from "IT Times". The system can convert machine information into natural language, sort out documents, manuals, and spare parts lists, and generate solutions to help operators and maintenance staff understand error codes.

From industrial production to healthcare, from travel planning to daily commuting, with the arrival of the "Agent - first" era, the future of intelligent agents is full of infinite possibilities - and it's only just beginning.

Trend 2: The way companies show their "muscle" has changed, and models are Agents

At the opening ceremony of WAIC 2025, Yan Junjie, the founder and CEO of MiniMax, said in his speech that as the capabilities of large - scale models continue to evolve, AI is shifting from a "chatting tool" to a more creative and executive productivity tool. He emphasized that the future of AI is not to simply replace a certain human skill but to systematically enhance individual and social capabilities.

At this year's WAIC, technology giants and star unicorns that used to focus on large - scale models have all shown off their "Agents". AI is no longer content with being a "smart chatting partner"; it's ready to "get to work".

The AutoGLM model promoted by Zhipu AI this time demonstrates task - planning and screen - understanding capabilities. On - site demonstrations show that after receiving instructions from users, the model can independently complete multiple operations such as grabbing red envelopes, integrating Xiaohongshu information, interacting on Weibo, placing orders on Taobao, and writing reviews. It can also handle navigation and ticket - booking tasks, getting closer to the intelligent assistant scenario.

Trend 3: From "single - agent" to "multi - agent collaboration"

No matter how powerful a single Agent is, it can still be overwhelmed by the complexity of the real world. "Multi - Agent collaboration" has become the "golden key" to unlocking complex tasks at this year's WAIC. A reporter from "IT Times" saw at the Midu booth that with just one instruction, visitors could experience the boost in office efficiency brought by AI: use the "V Assistant" to query public opinion data; use "Model Power Connect" to output official documents that meet requirements by simply entering a title; use "Proofreading Connect" to quickly complete the review of error - sensitive content in multi - format files. These tasks, which used to take hours, can now be completed in minutes.

Fan Ling, the CEO of Tezign Technology, has keenly captured the core driving force behind this change: "Since this year, there has been a fundamental change in the Agent field. The most significant change is the shift from a 'tool mindset' to a 'collaborative partner mindset'. Compared with large - scale models, agents have the ability to make autonomous decisions. The common application method is multi - agent collaboration - through the division of labor among agents in different roles, diverse tasks can be handled."

Fan Ling pointed out that previously, for large - scale models, everyone was aiming for the "efficiency tool" positioning. However, starting from GPT - 4o and after the DeepSeek - R1, AI large - scale models have gained Reasoning capabilities. They can engage in self - gaming, and their thinking ability has become stronger, enabling them to plan tasks. Now, people are trying to connect them with various tools. After the task planning is done, they can call on various tools for actual operation and directly deliver results. Therefore, Multi - Agent will be a huge opportunity.

Trend 4: From SaaS to RaaS, results are more important than services

Compared with the large - scale model era, one of the most prominent features of the Agent era is the priority of delivering results. Compared with the previous trend of "showing off" the complexity of functions and technologies, now, results are the key.

At this year's WAIC, the industry clearly felt that "practicality" has replaced "sensationalism". This is an inevitable microcosm of the development of large - scale models. AI manufacturers are starting to pay more attention to whether they can truly meet user needs and quickly deliver valuable results.

The MiniMax Agent exhibited by MiniMax has the ability to understand requirements, break down tasks, and execute them in multiple threads. It supports full - stack development and can complete delivery tasks such as web pages and e - commerce applications on the platform. During the demonstration by the booth staff for reporters, the MiniMax Agent directly generated a framework for a side - scrolling adventure game that could be "launched". Facing the automatic generation of page after page of code on the screen, you can hardly believe that your initial request was just a text of less than 100 words: "Make a side - scrolling game where a little man moves to the right on a line..."

The staff revealed that it only takes about half an hour to generate such a small game. Whether it's a video or a PPT, once the model - as - Agent receives the requirements, it can deliver results during the chat, greatly improving work efficiency.

Gao Wei, the chief analyst and deputy dean of the research institute at Midu, observed at the WAIC site that the evolution of AI technology has entered a new stage, especially in the practical application of large - scale models and agents. From general functions to more specialized and vertical solutions, with the explosion of artificial intelligence technology, RaaS (Results as a Service) has become a new solution.

The characteristic of RaaS is that it shifts the focus from numerous functions to actual delivery results. Users no longer need to worry about "one has 100 functions and the other has 120 functions". The most important evaluation criterion is whether the system can quickly deliver a high - quality result. With the continuous development of AI technology, the role of agents in the RaaS era is becoming more and more prominent, providing more efficient and convenient solutions for enterprises and users.

Trend 5: From B - to C - end

Behind the excitement at this year's WAIC is the explosive growth of C - end products, which have taken center stage. After the capabilities of basic large - scale models reached a usable level, the floodgates of C - end product innovation were opened. Inside and outside the exhibition hall, the frequency of appearance of various To C AI products and Agents reached a new high.

Some industry insiders pointed out that if there were still many differences in the understanding of Agent before, the biggest consensus now is that Agent is the software - form product of AI applied to the Internet. In the PC era, it was websites; in the mobile era, it was apps; in the AI era, it's Agents.

During WAIC 2025, WPS Lingxi, named the "treasure of the exhibition hall", attracted a large number of visitors. WPS Lingxi supports users to have multi - round conversations in natural language and complete functions such as document creation, presentation generation, and voice assistance. Its operation process does not require complex steps or jumping to external applications and can accurately preserve format details such as mixed text and images and complex tables, making work easier.

Trend 6: Agent infrastructure is gradually improving

The "last mile" for agents to penetrate all industries depends on both a strong ecological "circle of friends" and powerful infrastructure capabilities. At WAIC, the infrastructure and ecosystem around Agents are gradually becoming more complete.

Supermicro, relying on the AI - in - One Partner Program, released ten industry practices of the FusionOne AI solution. By building an industry agent matrix, it solves industry application problems through ecological collaboration. Technologically, it horizontally integrates technologies such as cloud services, virtualization, large - scale models, inference engines, and RAG to achieve technological integration. In the industry field, it vertically penetrates industries such as healthcare, finance, and education to provide customized solutions, enabling computing power to accurately deliver value.

On July 28, Alibaba Cloud launched "Wuying AgentBay", vividly defining it as a "super - brain" for AI Agents. In essence, Wuying AgentBay is a ready - to - use cloud infrastructure for Agents. It directly connects the tasks assigned by AI to the high - performance cloud environment. Developers only need three lines of code to "move in", completely getting rid of the cumbersome local environment installation and configuration.

Seeing these trends, PPIO did not choose to "compete in applications" but instead built an AI agent platform. PPIO launched the first domestic Agentic AI infrastructure service platform, comprehensively deploying from the first Chinese Agent sandbox to model services and then to the AI agent platform. It provides full - stack Agent infrastructure services for developers and enterprises, reducing the technical threshold and development costs.

At the Transwarp Technology booth, the staff presented a highly impactful technology demonstration, "Building an AI industrial Q&A intelligent agent in 10 minutes", compressing what used to take weeks of professional development into the time it takes to have a cup of coffee.

The Dark Side of the Moon booth is the most low - key among the "Six Rising Stars" in AI. It exhibited its main model, Kimi K2, which focuses on code capabilities and Agentic task processing and is more positioned as an efficiency - tool model for developers.

As the excitement of WAIC 2025 fades, those vivid images still linger in our minds: the robotic arms dancing to the code in the Siemens factory, the AI pre - diagnostician patiently examining patients, the industry agent "created" by Transwarp engineers in 10 minutes, and the office assistant that responds within minutes to a single instruction from an ordinary visitor... Together, they paint a clear picture of the Agent revolution: AI is transforming from a technological marvel on display to an indispensable "digital infrastructure" for social operation.

In the future, when developing an Agent becomes as easy as assembling Lego bricks and the agent - collaboration network penetrates every corner of society, we will no longer debate whether it's a passing fad or a long - term trend.

This article is from the WeChat public account “IT Times” (ID: vittimes), written by Jia Tianrong and Hao Junhui, and published by 36Kr with permission.