She Zongming/Wen As the two leading AI (artificial intelligence) applications in China, Doubao and Qianwen announced on July 4th that they will discontinue their intelligent agent functions on July 15th.

In the past two years, "intelligent agents" have become a hot topic in the field of AI. After OpenAI launched GPTs in November 2023, top AI assistants in China followed suit, with platforms such as Doubao, Qianwen, and Wenxin opening up entry points for custom creation of intelligent agents. At that time, the common saying was that big models were the "brain" and intelligent agents were the "hands and feet" - with the support of exclusive role positioning, private knowledge base, and API tool calling capabilities, AI assistants could evolve from "chatbots" to "clerks".

Intelligent agents have indeed served as an important tool for multiple platforms to attract new customers and retain them. Relying on personalized roles and scenario based interactions, massive intelligent agents were once the main incremental contributors to platform traffic.

Since that's the case, why do Doubao and Qianwen have a tacit understanding of the so-called "intelligent agent function" and raise a red light?

It needs to be clarified that this adjustment does not mean the demise of intelligent agent technology itself, nor does it mean that the two platforms have to give up the intelligent agent track. Only anthropomorphic interactive services such as user generated intelligent agents and custom roles have been cut, and the platform's official built-in standardized Q&A, office, and creative tool functions have not been affected.

On the timeline, it is not a coincidence that Doubao and Qianwen chose July 15th as the offline node, but rather precisely aligned with the official implementation date of the country's first AI anthropomorphic interaction special regulation, the "Interim Measures for the Management of Artificial Intelligence anthropomorphic Interactive Services". The new regulations, combined with the previous Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents, form a dual regulatory framework, drawing a red line for open self built intelligent agents within the general large model.

Prior to this, the intelligent agent creation function opened by Doubao and Qianwen aimed to create a UGC intelligent agent (self built intelligent agent) ecosystem: ordinary users only need to input prompt words and set character personalitiesAndDialogue memory logic enables one click generation of exclusive AI characters. Due to the low threshold, it quickly emerged on the platformGo OutMillions of users have built their own intelligent agents, which include learning assistants, workplace tools, as well as high-risk types such as virtual lovers, vulgar role-playing, celebrity impersonation, tarot divination, and gambling.

The new regulations impose strict regulatory requirements on services that simulate the personality traits of natural persons and engage in continuous emotional interactions, including mandatory indicators such as role registration, real-time review of the entire process, graded control of minors, and risk tracing. Considering the enormous audit pressure and high compliance costs, eliminating risks from the source has become the best option for Doubao Qianwen.

Setting aside policy factors, the adjustments made by Doubao and Qianwen are also an inevitable choice for the AI industry to shift from the expansion period of burning money for users to the verification period of business models - it should be noted that user built intelligent agents are already inefficient sectors with serious input-output imbalances.

From the perspective of computing power cost, user built intelligent agents mainly rely on fragmented high-frequency chatting and emotional companionship. Although a single interaction may seem lightweight, the computing power and storage consumption under massive calls are extremely astonishing.

From a commercial perspective, the conversion rate of payment for such pan entertainment scenes is extremely low, making it difficult to form a stable commercial closed loop. Although intelligent agents are a powerful tool for attracting traffic, the fact shows that the retention quality of the "entertainment demand satisfaction" users they attract is extremely low, making it difficult to convert them into platform paid members and high-value productivity scenario users.

For Doubao and Qianwen, after the user base is relatively stable, the urgency of relying on pan entertainment intelligent agents to exchange traffic is greatly reduced. Against the backdrop of industry wide efforts to reduce losses and control cloud costs, UGC intelligent agent businesses that consume only resources and have no stable returns have naturally become a priority for reduction.

The simultaneous offline of Doubao and Qianwen UGC based intelligent agents also marks the formal completion of the stratification of domestic AI product forms - the future AI application ecosystem will present a clear differentiation pattern: comprehensive tool AI strictly controls UGC anthropomorphic interaction, while anthropomorphic emotional interaction is separately operated by vertical product filing.

This is already evident in the response strategies of both platforms. ByteDance migrated relevant businesses to its "Cat Box App", trying to build an independent risk control and juvenile protection system in this vertical product; Alibaba's Tongyi Qianwen has adopted a more cautious strategy by divesting its compliance intelligent agent business from the main site and independently planning its launch and operation in the future. The physical isolation of "using a universal main station as a tool and vertical products as companions" can not only effectively achieve risk isolation, but also enable different scenarios to adapt to different audit standards and interaction logic, greatly improving governance efficiency.

For C-end users, this move is bound to bring pain in the short term. Many users who are accustomed to being accompanied by specific virtual roles need to migrate data, change products, and rebuild roles, which means significant switching costs and reminds users to establish a sense of data sovereignty and make good use of the window period provided by the platform for backup and export.

In the long run, this adjustment will force industry resources to concentrate in the direction of truly possessing social value and commercial potential. The research and development focus of underlying intelligent agent technology may accelerate its shift towards high barrier tracks such as enterprise level intelligent agents, industry-specific intelligent agents, AI workflows, and government enterprise services. A more secure, trustworthy, and sustainable AI ecosystem is far more vibrant than a square filled with vulgar characters and fake traffic.

The shift of Dou Bao and Qian Wen marks the official end of the extensive model of "everyone can play an AI role". Intelligent agents will not disappear as a result, they have simply shed their entertainment oriented facade and returned to their essence of empowering various industries. From this perspective, instead of seeing it as a regression of the AI model, it is better to see it as a painful but inevitable "molting".

(The author is a senior media professional)