Behind the rise of Agent in 2024 is not only the product of technological boom, but also the beginning of the fall of big models from the altar.

The halo of large models is gradually dissipating.

"In 2023, everyone generally thinks that they need to buy a large model, and many people are keen to train their own large models. But after training, everyone is confused about how to use it." The essence is that the large model presents The underlying capabilities, like water, electricity and coal, are “not products”.

The definition of "using large models" by the above-mentioned industry insiders is that it cannot stop at opening ChatGPT and Kimi to ask questions, but AI can reduce costs and increase efficiency for enterprises.

If the theme of the previous year was "the explosion of large models", the theme of last year was "the implementation of large models". People have gradually realized that it is not enough to rely on the power of the model alone, and how to use it is the "stuck" problem. Agent plays an intermediary role in connecting large models and scenarios. In particular, the AI ​​workflow truly allows the industry to see the path to implementation.

Some people think that "Agent mainly solves some of the long-standing difficulties in B-side delivery", while others comment that "model companies choose Agent for external sales solutions."

Agent makes large models in the cloud "down to earth", and "large model + Agent" has become the new fashion of To B.

LinkAI co-founder & COO Su Chenxing has a strong sense of the popularity of the Agent industry. “About June 2023, when we first entered the market, there were only three domestic companies making Agent tools including us. It was not until the second half of last year that the number began to increase, and there was even a trend of oversupply."

According to Qichacha, within 1-3 years, there will be more than 200 newly established companies related to intelligence.

The term "chaos has begun" accurately describes the current situation of the entire Agent market. Player identities are mixed, and big manufacturers, AI tigers, and start-up companies have all come to an end. Without the establishment of standards, the overall span of the industry is particularly large, highlighted by the unit price per customer, which ranges from several thousand yuan to tens of millions of yuan.

“At this stage, Agent is still a pseudo-concept, an intermediate product, and the real Agentic AI is full of imagination.”

At the end of the year, the person in charge of the 2B business of a certain company was preparing to price the upcoming new Agent product. He revealed that it was "quite cheap" and the price directly beat the big manufacturers. "The algorithm is 3,500 yuan and the research and development is 2,500 yuan, both of which are 20% off." .

Such a situation is extremely common in the current Agent market. Unlike the "clear price tag" of large models, the price system of Agent is very confusing. According to Su Chenxing, the unit price of Agents ranges from several thousand yuan to tens of millions of yuan, and the spectrum corresponds to scattered orders received by individual developers to large-scale government and enterprise projects.one.

Starting from the supply and demand relationship at both ends, it may be easier to understand the current development status of domestic Agents. Here we only discuss the delivery of Agents on the To B side. Demand drives the rapid implementation of Agents. There are two main lines here: top-down and bottom-up.

Su Chenxing believes that the completely opposite decision-making link determines the agent's unit price. Top-down generally corresponds to government and enterprise customers. Their logic is driven by intelligent transformation tasks, and they tend to purchase first and then sort out which scenarios and businesses they want to implement. Because it has to meet hard targets, the budget is relatively loose, and projects can often be worth millions, which opens up the upper limit of the Agent market.

In the bottom-up decision-making chain, the key figures are those who "eat crabs" in small and medium-sized enterprises. They have accumulated a certain amount of technical and practical experience, and clearly understand the effects that the Agent will have when implemented, and it is up to them to decide whether and which AI products to purchase to solve problems in business scenarios. Because they put their needs in advance and purchase with awareness of problems, the budgets of this group of people are generally not very high, but due to the large volume, they still account for most of the Agent market share.

Demand determines supply, and it can be divided into four tiers based on its ability to handle demand. At the top of the pyramid are major manufacturers such as Baidu, Alibaba, Huoshan, and Tencent. Cloud model infrastructure is the most complete and the amount of orders is the largest. At the bottom are manufacturers such as iFlytek and Zhipu. Their infrastructure is relatively complete, but they are struggling to accept some very large government and enterprise customers. The third level is transformed or emerging agent companies such as Lanma, Real Intelligence, Dify, and LinkAI, which focus on serving small and medium-sized enterprise customers; the last level is individual developers, who can use tools to meet some simple needs.

Compared with the CV era of the four AI tigers, large models have made To B’s life a little easier, and the average unit price of projects has increased from hundreds of thousands to several million. From the perspective of business personnel, Agent solves the pain point of difficulty in B-side delivery in the CV era. In the past, countless small models were superimposed to serve one business scenario. Now it is "large model + small model". The two are a combination of generalization ability and accuracy. Agent plays a connecting role in it, quickly and efficiently in the form of low-code projects. Build an application for customers to use.

However, as time goes by, the so-called "big orders" have become less and less. Some people in the industry reported to Photon Planet that in the early days, a To B project of Zhipu could easily be reported for more than 10 million yuan. The content was roughly model plus fine-tuning, but today it is no longer possible to report.

The "big order" wrapped in Agent's coat looks like a software application, but in fact it is a package of solutions, including cloud services, models, hardware, tool layers, industry knowledge and after-sales services. This means that competition in the large order market has an implicit threshold. When customers are “not short of money”, whoever has better infrastructure and more mature service system will be more competitive. The final winner is basically locked in the large order market. In the middle of the factory.

Six Little Tigers’ Zhongzhipu was obtained on the B sideThe most fierce, official data revealed that the total number of large model contracts in 2023 is 350 million yuan. Its vice president Chen Xuesong has worked in Alibaba Cloud and Megvii and has rich To B business experience. According to industry insiders, Zhipu has decided on the B-end direction from the beginning and recruited a large number of former Megvii employees to develop the B-end market for large models. The route also continues the idea of ​​"integrating software and hardware" from the previous era.

However, the innate frailty prevents Zhipu from breaking through the second-tier barrier. Many people told Photon Planet that "Zhipu plans to go public in the future, and there is heavy pressure for commercialization." The above-mentioned people revealed that Zhipu’s latest focus has shifted to Xinchuang and actively adapted to Huawei H920B.

“The investment cost of Xinchuang is not low, but other AI manufacturers have not done it yet. It is an option for Zhipu to choose to win the order with Huawei.”

The position of large models has become clear, and there are only a few manufacturers with the ability to provide models. Players in both the first and second echelons are essentially using Agent as a starting point to develop the cloud and large model market. Agent is the appetizer, while cloud and large models are the main course.

Apart from the one to two hundred companies competing for it, the demand for Agent applications from the remaining millions of small and medium-sized enterprises in China has not yet been fully recognized. Su Chenxing believes that the large demand in the future can still support the current situation. The third echelon of Agent application and service companies.

LinkAI is a startup company that emerged along with Agent. It started as an open source project that connected large models to WeChat ecological conversations. The product has evolved from the initial conversation assistant to include multi-modal large model aggregation services. , RAG knowledge base and Chat BI database, plug-in tools, Chat Bot and workflow construction, etc. to build SaaS products with zero-code agents.

According to Su Chenxing, Agent products that serve small and medium-sized enterprises can be divided into two categories, one is purely tool-based products, and the other is products with business scenario attributes. The pure tool type has high versatility, relatively light team investment, and generally does not require a sales team; while the delivery of products with business scenario attributes is heavy, and it takes time to accumulate industry experience like a snowball. After weighing the situation, LinkAI made a choice, focusing on making general-purpose SaaS products, and leaving a small part to provide customized services in marketing, e-commerce and other scenarios.

LinkAI is a microcosm of this wave of small but beautiful Agent startups. On the one hand, it quickly generates commercial revenue through standardized product delivery, and on the other hand, it also leaves behind for it to connect with large manufacturers and B-side customers at the same time. space.

According to LinkAI’s disclosure, since commercialization began in early 2024, the subscription revenue of pure SaaS has exceeded 2 million ARR, 70% of which comes from the natural conversion of open source projects and the word-of-mouth spread of PLG; To B project system The confirmed revenue has exceeded one million yuan, and the orders are even more.

Recently, LinkAI started cooperation with Baidu. There are two ways to promote cooperation. One is to put it on the Baidu Cloud Application Market for Baidu's customers to use, or it can divert traffic to its own products to complete the conversion; the other is to cooperate with Baidu on the To B project and provide it with Agent. Tool layer and cross-channel capabilities.

Even if a large manufacturer has a dedicated platform for building intelligent agents, it still needs to cooperate with startups. The factors to consider are simple. Rather than coordinating across departments, it is more efficient and time-saving to directly introduce third-party cooperation. Agent only accounts for a small part of large orders, and the input and output are not directly proportional, but startups can achieve relatively good results with tens of thousands of dollars. The cross-ecological flexibility of startups is also one of the advantages that big companies value.

Because of this, in the minds of Agent start-ups, the boundaries between business and profit are naturally drawn from those of large companies. Things don't seem to be developing on the established track, and the biggest uncertainty comes from the Byte Volcano Engine.

In the last press conference, Volcano Engine focused on introducing HiAgent, a product for enterprises to develop large model applications and agents. Its definition of Agent application construction is roughly the same as that of startup companies on the market, and can be regarded as a direct competitor.

The official explanation from Zhang Xin, Vice President of Volcano Engine, "If the beanbao model is compared to Android, then HiAgent is the SDK (software development kit) for enterprise scheduling system capability development applications."

According to insiders close to Volcano Engine, their internal assessment indicators have changed. "In addition to whether the bean bag model is used after it is deployed by the customer, it also depends on how many Agent scenarios have been implemented."

The former Button has cast a shadow. It was well-received but not popular. The C-side cannot bring in revenue. After the team was laid off, the To B business line that was merged into the Volcano Engine has not seen significant improvement. HiAgent is directly targeting corporate customers this time, and its competitiveness remains to be tested.

Now on the market The definitions of Agent are similar. For example, the Volcano Engine defines expert-level Agent applications as "private data + large model + Advanced RAG + Workflow".

When talking about the differences between Agents, many people mentioned a key word: industry attributes. Su Chenxing gave us an example. In the e-commerce customer service scenario, customers will first get a general product. Based on it, they will make fine-tuning for different products and copy relevant industry workflows with one click based on the template. There is also relevant training at the business level to guide customers in writing prompt words, building workflows, supporting them in importing relevant industry data, building knowledge bases, etc.

For now, intelligence is still a pseudo-concept,"The current product form is not the end, but the intermediate state of the entire industry."

The large model has expanded the agent's capability boundaries as never before, but from the definition, the current agent has no self-reflection and self-planning The ability is just to perform tasks according to the arranged process. Although it is generally believed that in the next 1-2 years, "large model + agent" will become the most popular model. The mainstream paradigm, but at this stage its essence is still a low-code product.

“In the past, when low-code was promoted, the public’s awareness and acceptance was not high. Now with the promotion and popularization of large models, people are more interested in Agent. The desire to explore and accept products has been significantly improved," Su Chenxing said.

The judgment of the intermediate state is also in line with the development direction of the entire technology path. The entire industry is moving from "AI Agent" to "Agentic The era of "AI" has changed, emphasizing the effect of executing single scene tasks to the comprehensive ability of independent planning, decision-making and execution of tasks.

It is a cliché to discuss SaaS by placing Agent in another dimension. Currently, in addition to projects system, the main business model of Agent is still subscription system

In foreign countries, Agent is SaaS. The financing and commercialization of the market have injected new vitality. The relaxed and friendly environment has also attracted some Chinese Agent companies to go overseas. It is reported that LinkAI is currently planning to launch overseas products at the end of this month.

In China, some traditional SaaS vendors are beginning to seek transformation, adding AI capabilities to their original products, and iteratively transforming to AI. SaaS creates a new dilemma: after superimposing AI costs, the unit price goes up, but it also weakens the relationship with some "AI The market competitiveness of "native" Agent products. Users prefer low-priced products with similar functions.

Su Chenxing said that for the SaaS market, AI may be half and half. Such as data analysis, customer service Such scenarios are very suitable for large models. This part of the market may be replaced by AI, but the remaining professional traditional SaaS still has an irreplaceable role.

“Large models cannot be seen, but Agent actually provides them. An opportunity for quick trial and error. If it is suitable for the business, quickly introduce transformation; if it is not suitable, stick to the original method. ”