From the “big model craze” to the “agent craze”, is it a “real trend” or a “false proposition”?

In 2023, the keywords of the AI ​​industry shifted from "big model" to "Agent" . Whether it is technology giants, startups, or government agencies, they are all frantically promoting Agent-related technologies. OpenAI's ChatGPT Plugins , Auto GPT, Microsoft's Copilot, Baidu's "Lingjing", Alibaba's "Qwen - Agent " , etc., all major manufacturers have entered the market. According to IDC ( International Data Corporation ) research , the annual compound growth rate of the global AI market is expected to climb to more than 20% by 2025.  

But at the same time, doubts continue to arise: Is the emergence of the Agent concept simply because "AI can no longer tell new stories"?

"Manus, please help me make a PowerPoint analysis of Tesla's 2024 annual report, about ten pages."

You only need to speak such instructions to the computer, and Manus will search for Tesla's relevant data and information online in an independent virtual machine, call office software, generate a complete PPT, and finally output the PPT to you in the form of a file. This is the function claimed by Manus, the world's first general-purpose AI intelligent product.

In the early morning of March 6, the domestic AI team Monica (Butterfly Effect Co., Ltd.) released the universal AI Agent product Manus in English on the foreign X platform, which instantly set off the AI ​​circle. First, a large number of self-media focusing on AI technology praised Manus, saying that Manus, as the world's first universal agent, has crushed OpenAI in terms of strength, and shouted the slogan "Manus makes Silicon Valley sleepless and ChatGPT silent." At the same time, the invitation code to experience Manus first was sold at a sky-high price of 100,000 on Xianyu .

Image source: Screenshot of a post published by ManusAI on the X platform

However, less than 48 hours after Manus became famous, public opinion took a dramatic turn.

As the popularity of Manus grew, many people began to question whether the functions of Manus' official demonstrations were technically innovative and were essentially "packaging" mature technologies into products. Coupled with the hunger marketing formed by the invitation code mechanism, some people even directly regarded Manus' popularity as a carefully planned hype to sell virtual currency by taking advantage of the popularity of AI.

From "another big bomb after DeepSeek" to falling into the whirlpool of public opinion of "shell + over-marketing", Manus's reputation staged a shocking reversal in just a few days. To this day, even though the official team has tried its best to clarify some controversial issues, the public opinion has not changed significantly.

According to Future Turing, after Manus appeared , the relevant departments once believed that it had typical value and quickly prepared to report it to the higher authorities. However, due to the continuous doubts about the originality and potential risks of Manus technology on the Internet at the same time, out of prudence, the department finally withdrew the application documents and postponed its promotion as a typical example.

In fact, since last year, major technology companies and research institutions have been frantically developing AI Agents. Baidu's Robin Li clearly stated last year that "Intelligent agents are the most mainstream form of AI applications and are about to reach their breaking point ." OpenAI CEO Altman also ranked the importance of intelligent agents at the forefront of the company's development strategy last year.  

However, is Agent really the only solution for AI implementation? Behind this wave of enthusiasm, has AI fallen into the "trap" of Agent?

Everyone is pushing Agent

In the field of AI, Agent is not a new concept. As early as 1995, computer scientist Yoav Shoham proposed the concept of software agent, which refers to a software entity that can run autonomously in a specific environment to achieve a set goal. However, before the emergence of large models, the capabilities of Agent were very limited due to the technical level, and the scope of application was relatively narrow. 

Until 2022, OpenAI released ChatGPT, which triggered a global AI boom. The powerful language understanding and generation capabilities of large models have injected new vitality into the development of agents. People began to imagine whether agents driven by large models can have the ability to think, plan and execute tasks independently, so as to truly become people's intelligent assistants.

In 2023, the emergence of experimental projects such as Auto GPT and Baby AGI has made people see this possibility. These projects are built on large models such as GPT-4, and can complete complex tasks such as creating and managing websites and formulating marketing plans by autonomously generating goals and plans and calling various tools (such as search engines, code execution environments, etc.) . Although these projects are still in a very early stage and their stability and reliability need to be improved, they have ignited people's enthusiasm for AI Agents.   

At one time, technology giants, startups, and research institutions all entered the market, trying to grab a place in this emerging field. Technology giants such as OpenAI, Google, Meta, and Microsoft increased their R&D investment in the field of AI Agents and launched a series of related products and research results. 

Domestic manufacturers are also making frequent moves. Alibaba not only launched the Qwen2.5-Max model, whose visual agent can automatically complete tasks such as query and reservation, but also continued to deepen its efforts in infrastructure and data services. On April 9, Alibaba Cloud announced that it would fully support the deployment and call of MCP (Model Context Protocol) services on the Bailian platform. Users can build an agent connected to the MCP service in 5 minutes, and will invest more than 380 billion yuan in cloud and AI hardware infrastructure construction in the next three years.

A research report by CITIC Securities predicts that 2025 will be the "first year of commercial explosion" for AI Agent, and the market size may reach tens of billions of dollars.

Looking back at the development of AI, every major breakthrough and change is inseparable from technological innovation and the expansion of application scenarios. From Turing test to deep learning, from expert system to GPT, the development of AI has always followed the dual-wheel drive model of technology-driven and demand-driven. For AI, the important thing is not to chase a hot concept, but to continue to invest in research and development, promote technological innovation and breakthroughs, and at the same time deeply explore user needs and expand AI application scenarios.

Agent may be an important direction for the development of AI, but it is by no means the only direction. In this field full of infinite possibilities, we need to maintain an open mind and innovative spirit to explore more possibilities and tell more wonderful stories.

Agent to Agentic 

With the evolution of technology, the concept of Agent is also evolving towards Agentic AI . 

On March 26, 2024 , Professor Andrew Ng first mentioned Agentic AI in a speech titled " Agentic Reasoning " at Sequoia Capital 's AI Ascent .    

Compared with traditional agents , Agentic AI is no longer a role that passively accepts instructions, but is endowed with the ability to actively think, plan and execute tasks, and is more like an intelligent entity that can make autonomous decisions. Traditional large-scale language model interactions are mostly for users to input instructions and the model to output results. This is a "non-agent workflow" and lacks iterative optimization space when dealing with complex tasks. Agentic AI uses the "agent workflow" to break down complex tasks into multiple steps, and gradually optimizes the results through cyclic iterations, which is closer to the human thinking mode of solving problems.  

Image source: Silicon-based technology

The OpenAI white paper emphasizes the " Agenticness " of Agentic AI , that is, the intelligence of the system, which enables it to flexibly adapt in complex environments and steadily move towards long-term goals. This means that Agentic AI is no longer just a passive data processing tool, but more like a capable partner of humans in the field of intelligence, able to reason, learn, and continuously optimize its own behavior like humans.  

Agentic AI is not only a data processing tool, but also an intelligent partner that can reason, learn and optimize behavior like humans. In the fields of customer service and process management, it demonstrates excellent efficiency and flexibility, autonomously handles customer needs and optimizes service experience. Its core concept is to empower machines with agency, so that they can proactively plan, predict and make decisions, and become a powerful assistant in business operations. Agentic AI marks the transformation of AI from an auxiliary tool to an autonomous manager .  

As computer scientist Alan Kay said , “The best way to predict the future is to create it.” As AI evolves from Agent to Agentic AI, we are witnessing the practice of this motto .

The evolution of technology has never been a black-and-white choice. Agent and Agentic AI are more like important coordinates in the AI ​​development map, rather than the ultimate answer.