Image source: Generated by Unbounded AI

"This year, we are building large models for specific scenarios, and we will conduct further exploration next year. We are currently discussing the issue of building digital employees with our cloud service partners." A logistics fulfillment platform unicorn told Industrialist.

If AGI is the ultimate goal of large models, then Agent is the key node in the process of achieving this goal. It is also the key to measuring the "from quantitative change to qualitative change" of large language models. In other words, only when the large language model reaches a certain level of maturity will Agent intelligence experience a real explosion.

In July 2024, OpenAI defined five levels leading to AGI (see the figure below), in which the Agent agent is located at L3, and claimed that we are about to reach the L2 level, which is the level of human reasoning. , can solve a variety of complex problems.

A few months later, Wisdom AI, the enterprise that foreign media called "the most likely to become OpenAI in China", expressed a different view when it released the intelligent agent AutoGLM. According to Wisdom, the large model has reached L3, which is the level where it can use tools and perform actions. However, its ability to master tools is not enough and it cannot learn by itself.

However, from the perspective of market demand, the trend of enterprise-level Agents has emerged. According to statistics, by 2028, the global Agent market is expected to reach US$285 billion.

For enterprises, the real value of AI is to reduce costs and increase efficiency, but current chatbots are far from meeting demand. Therefore, in the year of commercialization of large models in 2024, large enterprises, led by state-owned enterprises, began to develop industry large models to upgrade their internal IT construction, or to solve problems that are difficult to solve manually in specific scenarios.

But the problem is that not all companies have the "capital" to build large industry models or do work related to large model development; and after a year of exploring the commercialization of large models, one conclusion is that companies have It is not clear enough how to build large models and how to use them.

They need a set of "standard answers".

Unlike the large model, Agent is like this "standard answer", which is placed directly in front of the enterprise. Even the instruction manual and the return on investment ratio that the enterprise values ​​most are calculated.

The prelude to the “Agent War” has begun

A real data is that according to the foreign media Medium, by the end of 2024, there will be 500 million Agents are distributed in various industries. By 2025, this number will rise to 50 to 100 billion.

Overseas AI companies have joined the Agent war.

First, the AI ​​giant OpenA, which has raised more than US$13 billion, has joined the Agent competition; secondly, Anthro, which has raised more than US$7.3 billion.pic; AI Agent company Adept invested US$413 million to develop Agent; AI Agent company Imbue invested US$220 million; Magic AI spent US$145 million on Agent research and development.

In China, cloud vendors, large model vendors, operators, software vendors, etc. have already begun to explore Agent.

Starting from the beginning of 2024, Internet manufacturers led by Baidu, Tencent, and Alibaba have all released their own Agent development platforms. They are all based on their own large model platforms and launch low-code and no-code Agent development platforms for users. In addition, while launching the Agent development platform, they also launched a complete computing layer and model layer services for this purpose.

The Agent development platforms launched by the above-mentioned major Internet companies are, on the one hand, to expand the model ecosystem; on the other hand, they are also a new wave of competition for users in the era of large AI models. But in fact, strictly speaking, the Agents built through these development platforms cannot be called "Agents" as described by OpenAI. The latter is more inclined to the action level.

For example, playing the role of "digital employee" in an enterprise can truly reduce costs and increase efficiency.

As for agents that are truly at the action level, they are still limited to large enterprises. Among the large-model winning projects previously counted by industrialists, the results show that by 2024, smart bodies have become a trend, and in the industry distribution, telecom operators ranked among the top three in terms of the proportion of large-scale models purchased, mainly purchasing smart phones. The body is intelligent customer service.

In addition, in addition to cloud vendors and large model vendors, some software vendors are also trying to build intelligent agents using SaaS+AI.

A successful case is that overseas SaaS giant Salesforce launched SDR (Sales Development Representative) and Einstein Coach. Specifically, they can help companies screen and identify sales leads, arrange meetings, and then provide services to potential customers. Video images of similar-looking customers to help salespeople rehearse their words through role-playing, etc.

So why is Agent said to be the main narrative in 2025?

Because as large model technology, products and commercialization become more mature, what customers are most concerned about is not the list of large models, nor the new technologies of large models, nor the model architecture, etc. What enterprises really care about is Who can provide them with a standard answer, and who can match the enterprise's pain points seamlessly like a puzzle.

And these answers undoubtedly point to Agent.

2. From technology to implementation, Agent intelligence becomes the first step in AI

According to the large model winning projects in 2024, the main types of winning bids are Divided into three categories: computing power, industry large models, and intelligent agents.

NormallyIn this situation, only governments, large central state-owned enterprises, or industries that consume a large amount of GPUs, such as autonomous driving companies, operators, etc., have the need to purchase computing power.

Secondly, for large industry models, the purchaser is usually a large enterprise. On the one hand, developing large industry models also requires strong enough IT accumulation; on the other hand, building large industry models requires reorganizing the knowledge within the enterprise, and also involves breaking down the barriers between various IT systems, etc. This further increases development difficulty.

Moreover, judging from the commercialization exploration of large models in the past year, one conclusion is that companies are not clear enough about how to build large models and how to use them. Therefore, industry large models are not the most perfect option under certain circumstances.

But the intelligent agent is different. It is more like a standard answer, because the agent is more like an AI solution for a specific scenario in the era of large models.

For example, the intelligent customer service mentioned above is also the most widely used field today, and its value is obvious. The person in charge of a customer service large model project told industry experts that in the past, the resolution rate of intelligent customer service in the industry could reach about 70%, which means that the conversion rate to manual work was around 30%; but after applying the large model customer service, the resolution rate can be increased to 90% The above, for enterprises, saves tens of thousands of yuan in costs in 10 days.

This is the most real cost reduction and efficiency increase for enterprises.

Of course, intelligent agents are not so mature. First of all, from the perspective of industry distribution, the most widely used agents include intelligent customer service and AI code assistants. In this regard, industrialists have learned that among many Internet manufacturers, intelligent customer service is the first enterprise-level agent project they try.

Secondly, from the perspective of customer types, large enterprises are still the most willing to purchase.

You must know that in the era of large models, one of the most obvious changes on the demand side is the consumption of underlying resources, from CPUs in the past to GPUs. This means that companies have to consume more resources and require a lot of cost investment. Nowadays, both those who can afford large models and those who can afford intelligence are distributed in large enterprises/state-owned enterprises.

The person in charge of Baidu Smart Cloud Keyue revealed to industrialists that the biggest change in POC projects in the past two years is that the proportion of customers is more focused on large enterprises.

One thing to explain here is that the birth of any new thing requires exploration and innovation. Similarly, in the software industry, if a company needs to launch a new project, it usually has to go through a very critical and time-consuming step, that is, POC.

It is understood that some large manufacturers have already begun to cooperate with central state-owned enterprises on POC projects in smart bodies as early as mid-to-late 2023. In the future, as the intelligent agent ecosystem matures, these trends will also extend to small and medium-sized enterprises.

Who can get the Agent ticket?

In fact, many agents existed in the form of SaaS in the past, but now Agent is becoming a priority for enterprises.

In the era of big models, the transition from SaaS to Agent also means the subversion of the underlying architecture. In the past, the underlying architecture of SaaS was based on IaaS+PaaS; today, the underlying architecture is based on large models, that is, computing power layer + MaaS/model layer.

With this subversion of the underlying architecture, not all companies can get tickets for Agent.

Because Agent is designed based on large or small models, this means that Agent companies need to have model capabilities or cooperate with large model manufacturers. Take Real Intelligence as an example. It was once a traditional software manufacturer that mainly provided RPA solutions to customers. However, starting in 2023, it successively released its self-developed large models and began to transform into Agent.

Internet manufacturers such as Baidu and Tencent also need to rely on large model capabilities. It is understood that both parties have launched their own large model customer service robots, and the underlying model capabilities are also based on Wenxin Large Model and Wenxin Large Model respectively. Hunyuan large model. On this basis, fine-tune the model.

However, similar to the cloud computing era, the large model era also has more standardized Agent versions. These agents also exist in more standardized SaaS versions.

When a large number of standardized versions flood into the market, it is also the beginning of the Agent narrative in 2025.

So, apart from the underlying architecture, what are the differences between Agent intelligence in today's large model era and traditional software in the past?

One of the most obvious differences is that Agent is a software with self-learning ability.

Although at the current stage, large models have not yet developed to the extent that the Agent can fully learn and evolve on its own, which is what OpenAI calls the L3 stage; however, large model manufacturers rely on their past industry knowledge of serving enterprises. -how, summarize it into a SOP process and feed it to the Agent, and it can also evolve semi-autonomously.

In the future, with the further improvement of large model capabilities, Agent will reach the true self-learning stage. By then, more and more small and medium-sized enterprises will join the Agent narrative.

However, regarding enterprise-level Agents, or intelligence, there is no standardized paradigm from product route to internal AI construction within the enterprise, and from business model to service model.

Take the business model as an example. In the past cloud computing era, the payment models for SaaS software were mainly divided into subscription fees and customized development. In the future, from cloud computing to the large model era, when SaaS transitions to In the Agent form, more diversified payment models have emerged.

At present, it is mainly divided into three types:Category:

1) Billed according to the traditional SaaS subscription method; 2) Pay by tokens, which is also a new business model derived from the big model era, that is, pay on demand, and also pay according to Pay for the ability to call its Agent; 3) Through ecological cooperation, share according to actual results, such as sales growth, efficiency improvement, etc.; or cooperate with a system integrator to integrate the Agent into its products or services, through Achieve profits through sales sharing, cooperative promotion expenses, etc.

It can be seen that the narrative of everything about Agent is unfolding now.