Is AI More Than Just Large Models? Why Will Agents Become the Hottest Trend in 2025? | A Dialogue with Liu Zhiyi

36Kr's "Workplace Bonus" (ID: ZhiChangHongLi)

In the past decade, tens of millions of Internet products have emerged in China, but only 44 of them still have over 100 million monthly active users. Nine of these were launched in the past decade, and more than half of them belong to ByteDance, while two belong to Tencent [1].

Previously, almost no one could shake ByteDance's "dominance" in To C products. However, during the Spring Festival in 2025, DeepSeek became the first application software in history to reach over 100 million downloads within a month of its launch without any marketing expenses.

Before this, ByteDance was involved in a rumor that it had spent over 100 million yuan on monthly advertising. In this non - confrontational battle, DeepSeek [2] broke the decade - long "dominance" of well - known large companies, bringing confidence to industry practitioners.

However, in the field of foundation models where tech giants stand firm, there is no longer room for new startups.

For the continuous development of AI, the market needs new stories.

After the smoke of battle clears, everything begins to rejuvenate. The improvement of the large - model inference ability brought by DeepSeek has promoted the development of AI Agents represented by Manus. The enhancement of inference ability means that large models can understand users' intentions more deeply, improving the accuracy of decision - making and content generation. According to public data, within just two days of its release, the number of people who pre - registered for Manus exceeded 1.2 million, and it later increased to 2.6 million. If it weren't for the fact that running a single task on Manus costs 2 US dollars, this number would have continued to rise.

Manus has verified the market's enthusiasm for Agents. After Manus, competitors such as Baidu Xinxiang, Genspark, Fellou, and Coze Space have emerged one after another, and AI Agents have become the next direction recognized by companies.

Agents have become the hot topic that companies are competing for in 2025.

Currently, existing Agents have some drawbacks, such as slow response speed, a tendency to generate hallucinations, and unfriendly text - only interactions. However, their advantages outweigh the disadvantages: Agents not only have a smart large - model "brain" that can autonomously execute tasks and improve efficiency but also can help large models "grow hands and feet" to explore the world and even generate a scale effect. Through engineering technology, models can be equipped with the ability to continuously reflect and learn - which is impossible to achieve with the existing model structure.

On the other hand, the market has also voted with real money: Manus completed a 75 - million - dollar financing round in April, and its valuation skyrocketed five times. At the same time, industry giants have also voiced their opinions. Sam Altman, the CEO of OpenAI, said that in 2025, Agents will join the workforce; Satya Nadella, the CEO of Microsoft, predicted that Agents will replace some knowledge - based jobs.

When tech giants and the capital market jointly bet on the future of Agents, "Workplace Bonus" invited Liu Zhiyi, the author of "The Age of Agents", to conduct an in - depth analysis: What are Agents? How will the boundaries of human work be reconstructed in the age of Agents? How will the core value of humans change in a future where Agents take over repetitive work?

Liu Zhiyi is a leading Chinese AI scientist. He is currently a researcher at the Qingyuan Research Institute of Shanghai Jiao Tong University and also serves as the chief AI scientist of a listed company. For more than a decade, he has been conducting in - depth research and practice in the field of AI. He has worked at AI companies such as Baidu and SenseTime, mainly researching intelligent computing, spatial intelligence, and super AI alignment. He was selected as one of the "Top Ten Influential Figures in AI" by Forbes China in 2024. He has also written several academic monographs and popular science books, such as "The Enlightenment of Intelligence", "Embodied Intelligence", and "Digital Economics".

"When standardized thinking becomes the norm, non - standardized content becomes more scarce." Liu Zhiyi believes that in the upcoming age of Agents, non - standardized professional traits, such as more diverse, complex, and atypical ones, will become the core competitive advantage of humans over AI.

[1] Data source: "The product paradigm of large companies is shaken, and the possibility of entrepreneurship returns."

[2] DeepSeek is the parent company of DeepSeek, with the full name "Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.". It was established on July 17, 2023, and was founded by Magic Square Quantitative, a well - known quantitative asset management giant. The company is headquartered in Hangzhou, Zhejiang Province.

What are Agents, and what is the value of humans in the age of Agents? ╱ One

What kind of future will Agents bring to the workplace? ╱ Two

 

What are Agents, and what is the value of humans in the age of Agents?

Workplace Bonus: Why is it said that DeepSeek has opened a new era of Agents?

 Liu Zhiyi : The concept of Agents is not new. Its theoretical framework was first systematically proposed by Marvin Minsky in his book "The Society of Mind" published in 1986. Of course, with the progress of the times, the concept of Agents has also been newly defined. The Agents we currently define in a narrow sense are tools that can autonomously form plans, call tools, and have multi - modal capabilities based on large models. It is also the technical logic that upgrades the capabilities of AI from passive response to active execution.

When we discuss Agents now, we usually mention DeepSeek or the entire era of large models. Currently, there are two logics for judging the development of AI: First, large models transform from passive tools to active execution systems, a process from passivity to initiative; the other is the collaboration between large models and humans. Initially, the capabilities of models may be inferior to those of humans, but they gradually surpass most people, indicating an upward trend in the intelligence level of models. DeepSeek happens to be a key factor in promoting these aspects.

Workplace Bonus: What is the relationship between Agents and digital humans? Has the time come for Agents to be combined with embodied intelligence?

 Liu Zhiyi : Digital humans emphasize more on multi - modal visual technology. From 2015 to 2022, many digital humans have emerged in the market, and there are also relatively mature applications for digital humans. In fact, Agents are a more fundamental technical capability. It can be understood that the development of Agent technology may promote the automation of digital humans or other software engineering aspects.

Workplace Bonus: Just two months after Manus, we saw the emergence of competitors such as Baidu Xinxiang, 360, Genspark, Fellou, and Coze Space, but few of them are truly open for application. What does the development of Agents mainly rely on, and what are the current bottlenecks?

 Liu Zhiyi : Agents rely on the large - model "brain" for planning and decision - making, that is, they rely on the progress of foundation models. The stronger the foundation model's capabilities, the stronger the automation and generalization abilities of Agents when performing complex tasks. The breakthrough in this regard was the improvement of inference ability brought by Kimi 1.5 and DeepSeek before the Spring Festival.

Another core difficulty for Agents lies in long - term, close dynamic retrieval and modal association. Take a medical Agent as an example. We hope that when it makes a diagnosis, it can remember the patient's symptoms and conduct retrieval based on these symptoms to carry out a series of tasks, rather than having a diagnosis or actions full of hallucinations.

Workplace Bonus: What relatively mature scenarios can you currently see for Agents?

 Liu Zhiyi : Currently, the collaborative framework for multi - Agents has not yet converged, so more attempts are being made in vertical scenarios. For example, many large manufacturers have started to use Agent - based customer service on a large scale; in the financial field, intelligent investment advisors and intelligent investment research have also tried to use Agents; in the medical field, there are also Agent - based doctors for auxiliary diagnosis and treatment.

Workplace Bonus: Currently, the approaches of companies developing Agents may vary. Generally speaking, are there different schools?

 Liu Zhiyi : The companies I've communicated with can be roughly divided into two categories. One is the "technology" category, and the other is the "scenario" category.

The technology - oriented companies provide technical frameworks. By using their frameworks, the cost of using Agents can be reduced. These companies strive to make this Agent framework a standardized tool for other companies to apply AI. Manus has publicly stated that the cost of running a single task is approximately 2 US dollars, mainly due to its reliance on the Claude 3.5 model API of Anthropic and high - computing - power consumption (the token volume for a single task reaches the million - level).

The other category is the companies that focus on vertical applications in fields such as food delivery, finance, and healthcare, as mentioned earlier.

There is also a type of company that people may not pay much attention to, which develops open protocols. Take the most commonly used social software as an example. If it doesn't have a food - delivery function, it can develop an open protocol that allows users to call third - party food - delivery platforms on its interface to achieve the function of ordering food, similar to Apple Intelligence of Apple.

Workplace Bonus: What technological and commercial breakthroughs may indicate the end - game for the Agent market?

 Liu Zhiyi : One is the improvement of the collaborative ability of cross - organizational Agents. This will enable the hierarchical architecture of multi - Agents to converge into a unified technical framework, and standard technical tools will be formed.

Second, it is the improvement of the interaction ability between Agents and the physical world. Just like the Omniverse physical simulation platform developed by NVIDIA and the related world - model generation technology, if they can reach a mature stage and be put into practical use, it will also become a key technological breakthrough node in the field of intelligent operation.

Third, the maturity of the open - source ecosystem is also important. The R1 launched by DeepSeek is actually a typical symbol of the open - source achievements based on models. In the future, the maturity of the technical framework of the open - source ecosystem based on Agents will also bring a ten - fold or hundred - fold increase in the market scale.

 

What kind of future will Agents bring to the workplace?

Workplace Bonus: In the future, which jobs may have a strong collaborative relationship with Agents?

 Liu Zhiyi : This can be understood from two aspects. On the one hand, people often think that AI will replace jobs with high repetition and low creativity, but in fact, it is more about humans and AI collaborating to replace the original work models. An experiment conducted by a classmate during our exchange at the University of Hong Kong found that accountants, lawyers, etc. who understand AI can complete their work more efficiently. Therefore, an individual's work value may be replaced by colleagues who understand AI, rather than being directly replaced by AI.

On the other hand, from the perspective of managers, enterprises hope to use Agent clusters to improve production efficiency and promote organizational transformation. This will transform middle - level managers into "cognitive architects", including system thinking, human - machine collaborative thinking, and optimizing goal management and team collaboration through AI tools. Just as the steam engine impacted the craftsman system, the age of Agents will give rise to the "super - individual" and Agent - cluster models, and new occupations such as AI trainers will emerge.

Workplace Bonus: Do you think there are any industries that will be deeply transformed or reshuffled by Agents first?

 Liu Zhiyi : We can see that industries such as healthcare, finance, manufacturing, and education have already carried out in - depth innovation. This also includes the intelligent government. For example, Futian District in Shenzhen has deployed 70 "AI digital employees" based on the DeepSeek large model, covering 240 scenarios such as document processing and people's livelihood services, with an accuracy rate of nearly 90% for a single response.

Workplace Bonus: Which industries do you think Agents can bring benefits to?

 Liu Zhiyi : Personally, I'm quite optimistic about two fields. The first is the cultural and tourism industry because it involves the most modalities. For example, the AI - related work of venues such as the Forbidden City or the Water Cube, as well as digital tour guides.

The second is related to the real - estate industry, but not the traditional scenario of selling houses. Many industrial real - estate companies use Agents to manage their investment promotion and intelligent assistants in the rental market.

Workplace Bonus: Will companies have different requirements for talent in recruitment and selection due to the development of Agents?

 Liu Zhiyi : The development of Agents will change people's views and perspectives on AI technology. AI has transformed from a tool to a "colleague" that can collaborate with you at work. People will think more about how to combine with Agents. There may even be a need for people to have certain development or secondary - transformation abilities to make Agents better cooperate with work.

On the other hand, the role of managers will also change. Previously, they managed people, but after the development of Agents, they will manage intelligent enterprises and people.

Workplace Bonus: Do you think the domestic market for AI Agents will be larger than that for SaaS?

 Liu Zhiyi : At least in the domestic market, it is possible for it to become very large. Since Agents are not yet mature, there is still a large development space in the market. Without a consensus, it is very likely to create a new incremental market.

Workplace Bonus: For students approaching the college - entrance - examination graduation season who want to enter the field of AI, what majors are recommended for undergraduate and postgraduate studies?

 Liu Zhiyi : First of all, if the goal is to become an AI scientist, it is recommended to choose majors such as artificial intelligence, computer science, or applied mathematics for undergraduate studies, and systematically study core courses in mathematics, algorithms, and machine learning. Then, pursue a doctoral degree (92% of researchers in top - tier AI institutions hold a doctorate), focus on cutting - edge fields such as neural - symbolic systems and reinforcement learning, and publish papers in top - tier conferences. However, AI is currently in a stage of rapid development and requires a large number of interdisciplinary talents. So, on the basis of a computer - science undergraduate degree, students can study other fields they are interested in, such as film and television special effects, music composition, and finance.

However, it must be admitted that since the development of AI technology, many innovations have come from enterprises. Therefore, it is not recommended that students stay in school for too long. In China, undergraduate - degree students are already "over - educated". At this time, students can look for opportunities in the market first. It is recommended to gain some practical experience before further studies.

There's no need to finish all your studies at once.

Workplace Bonus: Did you plan your career when you were in school?

 Liu Zhiyi : After graduating from school, I had a clear expectation for my career path: starting as a researcher, gradually being promoted to a project leader, then to a technical director, and even to a higher - level manager. However, reality often doesn't develop exactly as planned - industry changes, company adjustments, and personal encounters can all cause twists and turns in the original path.

A career is not a straight - upward line but more like an exploratory path full of unknowns. Changes in the market environment, iterations of technical directions, and even an unexpected project failure can all challenge past accumulations. However, this doesn't mean a "negation" of career value but is part of growth.

True career resilience doesn't lie in being able to follow the plan exactly but in how to adapt to changes, adjust directions, and re - position oneself in new opportunities. Whether it's shifting to a new technical field, transitioning from technology to management, entrepreneurship, or cross - industry integration, the possibilities for career development are far more abundant than the initial plan. The key is to maintain the ability to learn continuously and find one's unique value in the face of change.

The current workplace system tends to make employees more professional and standardized, but isn't standardization what AI is best at? So, I think people at this stage should strive to make themselves non - standardized talents.

Workplace