Is the AI Agent the biggest opportunity or just a bubble in 2025?

ChatGPT can not only "talk" but also start to "do".

Early in the morning of July 18th, Sam Altman, along with four OpenAI researchers, officially launched ChatGPT Agent, a general - purpose AI Agent, during a live broadcast. This move made OpenAI, which had been quiet for a long time, attract attention again.

In terms of functionality, ChatGPT Agent is not very different from existing Agent products on the market. It can create tables, search for information, and execute tasks. However, many people still regard it as another "game - changer" from OpenAI because it validates a new mainstream direction - the model is the Agent.

Since this year, more and more large - model companies have placed their bets on the concept of "the model is the Agent". New products launched by xAI, Kimi, Lingyiwanwu, and Alibaba Cloud are all more or less related to it.

What is "the model is the Agent"? Why are large - model companies so eager for it?

1

"The Model is the Agent"

The Only Way to AGI?

Before 2025, large models and Agents seemed to be on two different paths.

The usual explanation is that large language models are like knowledgeable assistants that can answer questions and generate text, but they lack initiative and execution ability. An "Agent" is more like a virtual employee with clear goals and independent thinking ability. It can not only understand requirements but also take the initiative to act, execute tasks, and even interact with the external environment.

Take Manus, the world's first general AI assistant that attracted wide attention last year, as an example. By calling multiple underlying models, it achieved a closed - loop ability from planning to execution and then to output. It's like integrating multiple tools into one intelligent body. Although it's a "patchwork", the "patching" level is quite good.

However, industry insiders pointed out at that time that the barriers for general intelligent agents were not high. Li Dahai, the co - founder and CEO of Mianbi Intelligence, once mentioned: "Whenever there is a qualitative leap in the large - model version, it often swallows up the entire application ecosystem supported by the previous generation of models."

In 2025, the path of "the model is the Agent" became more and more obvious.

The so - called "the model is the Agent" means that the large model itself becomes the core brain and driving engine of the Agent. Different from relying on complex workflow orchestration or external model integration in the past, AI will evolve from being simply "eloquent" to being "capable of doing". This transformation makes AI an intelligent assistant that can actually help users complete tasks.

Take ChatGPT Agent as an example. Different from Manus, it completes the entire process of skill invocation and task execution within a single model. In this process, users can see the AI's operation path on the virtual computer in real - time and experience the whole process from understanding requirements, selecting tools, executing operations to delivering results. This "observable and interactive" Agent form is OpenAI's unique innovation in the technical foundation.

Previously, the large model globally recognized for its strongest programming ability was Claude 4 series by Anthropic. It "set a new standard" in programming, reasoning, and Agent capabilities, and can handle complex and long - running tasks. In fact, many intelligent agents are just "wrappers" of Claude 4.

This also made Claude 4 the target of criticism. On July 9th, xAI, an artificial intelligence startup under Elon Musk, released a new - generation large model, Grok 4, including single - agent and multi - agent versions. It has functions such as tool use and real - time search, targeting Claude 4 Opus.

Many large - model companies in China have also made a shift.

At the end of last year, Mianbi Intelligence, in collaboration with Tsinghua University, released a new - generation active Agent interaction paradigm. Agents are no longer simple instruction executors but have been upgraded to intelligent assistants with "insight". Li Dahai believes that the model itself, rather than the workflow, is the future development direction of AI intelligent agents. "The model is the Agent, the model is the product, and the model is the interaction."

On July 11th this year, Kimi launched a new - generation base model, Kimi K2, after a six - month interval. According to the official introduction, Kimi K2 is a MoE architecture - based model with stronger coding ability and better at general Agent tasks. The model itself integrates the ability of autonomous decision - making and task execution and can solve complex tasks as an intelligent agent.

Early in the morning of July 23rd, Alibaba open - sourced a new Tongyi Qianwen AI programming large model, Qwen3 - Coder, which also emphasizes Agent capabilities. It is particularly good at solving long multi - step tasks. It can not only plan work autonomously from a global perspective, support Agents to call various tools for in - depth research, but also finally solve complex programming tasks.

2

"Hallucinations May Not Be a Bad Thing"

Why are these large - model companies all betting on "the model is the Agent"?

"The concept of 'the model is the Agent' reflects a fundamental change in our perception of AI. It is not only a change in the technical architecture but also a change in the human - machine collaboration relationship," Fan Ling, the founder and CEO of Tezign, told a reporter from "IT Times". In his view, the core value of "the model is the Agent" lies in that large language models can simulate real - user behavior. AI can not only answer questions but also actively build user profiles and drive the decision - making process.

For example, if all the stories of "Harry Potter" are input into a large language model, the large language model can become "Harry Potter" and even think like Harry Potter.

This change in logic will bring a significant change to the industry, that is, Agents need to shift from a "tool mindset" to a "collaborative partner mindset", which may even impact some previous perceptions of AI. "People used to worry about hallucinations, but in fact, hallucinations may be a good thing. When we need AI to think actively, in terms of reasoning, we need to shift from the current 'convergence - first' to 'divergence - first' to make AI's thinking more open as much as possible," Fan Ling believes. This path can be successful, but the key at present is to find the right application scenarios.

A research report from CICC points out that the large - scale foundation model remains the key to determining the upper limit of Agent capabilities. The programming and intelligent - agent capabilities of large models are the focus of competition among various manufacturers.

However, compared with the more difficult "the model is the Agent", C - end Agents have greater market potential. The multi - agent collaboration model similar to Manus, which realizes diversified task processing through the division of labor among intelligent agents in different roles, is more common at present.

After the release of ChatGPT Agent, Manus immediately released 10 actual - test cases, trying to prove that ChatGPT Agent is not superior in task closure and visual delivery through different scenario tasks such as financial modeling, life planning, itinerary arrangement, consumer shopping, and flight screening.

However, the current general consensus in the industry is that the application of general intelligent agents is still in the early stage, in the process of business - scenario exploration and technical verification.

Gartner also believes that there is still a "bubble" problem in this market. It is estimated that by the end of 2027, more than 40% of intelligent - agent projects will be cancelled.

Fan Ling agrees with this view. He believes that just like the development process of any new technology, the Agent field will also go through cycles of "over - expectation" and "rational return". "The key to survival lies in whether the Agent can solve core business problems, and at the same time, find a balance between technology and cost. In Tezign's Atypica.AI practice, we found that when AI can directly deliver results, users' willingness to pay significantly increases."

3

The Market Landscape Has Changed

AI technology is still advancing at an accelerating pace, but the reality is that since 2025, the market landscape for large - model companies has changed significantly.

At the beginning of the year, DeepSeek entered the market strongly with its low - cost and high - performance open - source model. The so - called "Six Little Tigers" that attracted wide attention last year - Zhipu AI, MiniMax, Kimi, Jieyue Xingchen, Baichuan Intelligence, and Lingyiwanwu - have all experienced varying degrees of changes in terms of financing, traffic, and market share. Except for Zhipu AI and Jieyue Xingchen, the other companies have not received any more financing since the second half of 2024.

Meanwhile, news such as user loss, high - level departures, and business - line cuts has also emerged. In particular, Lingyiwanwu and Baichuan Intelligence gave up training large base models early and carried out strategic contractions. Baichuan Intelligence focused on AI in healthcare, while Lingyiwanwu placed its key bets on the implementation of AI in industries.

Intelligent agents that can directly "take action" have become the second hottest trend in the global technology circle. Market consulting firm Gartner listed AI Agents as the top strategic technology trend in 2025.

Regarding how to implement AI Agents, different people have different approaches.

Source: unsplash

Early this year, a reporter from "IT Times" interviewed several industry experts about the future of Agents. Yang Guosheng, the vice - president of Fanwei, said in an interview with the reporter that when general large models are applied in vertical fields, their stability is not satisfactory. This means that in the B - end market, accuracy is particularly crucial. "A large amount of engineering technology needs to be involved to ensure that it can truly achieve the effect of enterprise - level applications."

Half a year has passed, and Agents are still in the spotlight, but it seems they are not the same as before.

"Since this year, there has been a fundamental change in the Agent field," Fan Ling pointed out. In Tezign's Atypica.AI practice, he found that many people still regard Agents as an upgraded version of traditional tools. In fact, the connotation of Agents has already changed. Especially after the reasoning ability of large models has been significantly improved, it can directly give execution results by combining multiple tool invocations. Multi - Agent will be a greater opportunity.

However, Fan Ling also believes that general Agents cannot solve all problems. When technology meets unmet needs, there is room for growth. The value of vertical Agents lies in their ability to solve specific industry pain points.

4

Convenience and Risks Coexist

Meanwhile, some risks cannot be underestimated.

From a technical perspective, AI models may face bias risks. For example, algorithmic discrimination may lead to unfair results, and insufficient training data or improper application may also cause model failure. In addition, network security issues cannot be ignored. DeepSeek once suffered a DDoS attack.

A reporter from "IT Times" also noticed that after the release of ChatGPT Agent, OpenAI immediately issued a long - warning, reminding users of the potential risks of using AI Agents. Although ChatGPT Agent performs strongly in handling complex tasks, OpenAI emphasizes that the potential risks of the product still exist. For example, criminals may try to "trick" AI agents into providing private information that should not be provided or taking inappropriate actions.

Source: unsplash

To this end, OpenAI has strengthened the control of high - risk tasks and introduced a number of security measures: Key operations must be explicitly authorized by the user; High - risk tasks (such as sending emails) require the "supervision mode" to be enabled, requiring the user to monitor the whole process; For high - risk instructions such as bank transfers, the AI will actively refuse to execute; Users can clear browsing data and log out of the session at any time or disable connectors when not connected to the Internet.

Chen Yunwen, the CEO of Daguan Data, also admitted that the popularization of AI Agents among the general public still faces bottlenecks. The general public has a limited understanding of its functions and value and has a low acceptance level. There is also a shortage of professional talents, especially those who are proficient in both technology, products, and are familiar with business and the ecosystem. At the same time, in terms of ethics, law, and security, AI decisions often lack interpretability, the responsibility attribution is unclear, and there are risks of privacy leakage and attacks.

Anyway, ChatGPT Agent still needs to make further progress in solving practical problems. AGI (Artificial General Intelligence) remains the much - anticipated "starry sea", and the repeatedly delayed GPT - 5 is "the hope of the whole village".

This article is from the WeChat official account "IT Times" (ID: vittimes). Author: Jia Tianrong, Editors: Qian Lifu, Sun Yan. Republished by 36Kr with permission.