Recently, Xiaohongshu's engineering team REDtech published an article officially unveiling a Personal Agent called "Mishu," positioning it as an "AI life companion."

Based on the application scenarios officially disclosed, Mishu is in no rush to showcase the complexity of its technical capabilities. Instead, it focuses on concrete everyday problems—finding hole-in-the-wall shops beloved by locals in an unfamiliar city, discussing how to order from a menu, comparing the real-world user experience of two products, catching up on taste-test reviews of new drinks, or discovering exhibitions and markets worth visiting over the weekend based on recent notes.

For over a decade, competition among internet platforms has revolved around one core question: how to get users to spend more time on their products.

Search engines compete for search demand, short-video platforms compete for user attention, e-commerce platforms compete for shopping behavior, and local-life platforms try to keep consumption needs—eating, drinking, and entertainment—within their own ecosystems.

Yet although the business models differ, the underlying logic is remarkably consistent: first acquire users, then—through content distribution, advertising, or transaction services—convert user attention into commercial revenue.

The emergence of AI Agents is gradually changing this logic.

When AI is no longer content to simply answer questions but begins to understand user intent, call external tools, coordinate different services, and even complete a series of operations on behalf of users, what internet platforms compete for changes accordingly.

Users may no longer need to open multiple apps to search for information, nor repeatedly compare across different platforms. A single Agent has the opportunity to complete the entire continuous process—from need identification and information screening to service invocation.

In other words, the traffic gateways, information distribution, and transaction links previously controlled separately by different internet platforms may be reorganized.

Since the second half of the year, such changes have been particularly evident. In September, Meta launched the personal agent Muse, and OpenAI subsequently released Dots; in the domestic market, Alibaba's Qwen Go and Ant Group's Lingguang 2.0 were unveiled one after another, while ByteDance and Tencent are also accelerating their related layouts.

AI is transforming from a tool that users must repeatedly open and actively query into an assistant that has the opportunity to stay by users' side for the long term.

Now, Xiaohongshu is entering the fray as well. It may look like it is simply riding the Personal Agent trend, but if Mishu is placed back into Xiaohongshu's business accumulation over the past decade or so, it becomes clear that its choice is somewhat different.

Alibaba has a transaction and service ecosystem, Tencent has social relationships and application gateways, and companies like OpenAI are exploring more general-purpose personal execution environments. Xiaohongshu's most valuable resource is the vast amount of consumption experience generated around specific life scenarios.

In the past, this experience helped users discover products, find restaurants, and plan trips; now, Xiaohongshu hopes to let AI directly call upon them and participate in users' real-world decisions.

Suppose a consumer is preparing to buy a camera.

In the past, he might first learn about the specifications on a search engine, then go to Xiaohongshu to read real reviews, then open e-commerce platforms to compare prices, discounts, and after-sales service, and finally place an order. Throughout the process, each platform solved part of the problem, but it was always the user who connected these services together.

Now, Agents hope to take over this work.

Users need only state their budget, usage scenarios, and preferences, and AI can attempt to complete product screening, information comparison, price inquiries, and even invoke transaction services. A process that originally required multiple applications has the opportunity to be compressed into a single interaction gateway.

This is also the biggest difference between Agents and traditional AI Q&A, but judging from the current layouts of various companies, Agents have not converged on a completely identical technical route.

Take Alibaba's Qwen Go as an example. Its advantage lies in its ability to call ecosystem resources such as Taobao, Amap, and Fliggy, further connecting user needs to transactions and services. In public tests, Qwen Go has demonstrated capabilities such as checking flight delays, contacting airlines to seek compensation, and locating Taobao price-protection entrances.

Ant Group's Lingguang 2.0 emphasizes the transaction-processing capabilities of a personal AI assistant, hoping to integrate application creation, tool invocation, and specific services into the same interaction process; Meta Muse and OpenAI Dots explore continuous service capabilities from the perspective of a long-term personal assistant.

Products from various companies are still at different stages of development, but the focus of competition has gradually extended from the quality of a single answer to how much personal information and action permission users are willing to hand over to AI.

Xiaohongshu's entry point is somewhat different. The platform itself does not have a complete business ecosystem covering shopping, travel, and payments like Alibaba, nor does it need to replicate all the capabilities of a general-purpose Agent at the first step. Where Xiaohongshu can more easily leverage its advantages is instead those consumption questions that have no standard answers but are highly dependent on life experience.

Take restaurant recommendations as an example. Ratings and menus can tell users the basics of a restaurant, but "whether it's suitable to take parents there," "which dish is worth ordering," and "whether it's easy to queue on a weekend evening" often require more specific consumption experience to answer.

Such questions are difficult to solve simply by relying on product specifications, map ratings, or search results. Users also need to refer to other people's experiences and make judgments based on their own budget, taste, time, and specific scenarios.

This is also the content value that Xiaohongshu has accumulated over the years. The Mishu introduced by REDtech has already incorporated these scenarios into the scope of product exploration. For example, in an unfamiliar city, users can ask which nearby hole-in-the-wall shops are worth visiting; when they reach the entrance of a restaurant, they can learn about relevant reviews; and after entering the store, they can ask about dish pairings while looking at the menu.

One particularly interesting technology among them is what Xiaohongshu calls "progressive perception."

Traditional search often requires users to describe the problem clearly all at once, whereas in real life, people's needs often change as the scenario evolves: when just entering a restaurant, users may only want to know how the place is; after seeing the menu, they then further consider the number of diners, taste preferences, and portion sizes.

What "Mishu" does is try to gradually acquire on-site information as the task progresses, combine it with previous conversations and personal background, and offer suggestions in a timely manner. With the additional support of first-person interaction through Rokid AI glasses, users can ask while looking and do not need to re-describe the environment every time.

In other words, search engines solve where to find information, community platforms accumulate how others have experienced something, and what "Mishu" wants to do is place this experience into a person's specific situation.

Community experience can be said to be Xiaohongshu's most visible advantage in building a Personal Agent.

However, there is still a distance from content to decision-making. Xiaohongshu has a large number of user notes, covering life scenarios such as dining, travel, outfits, home, digital products, and beauty. Compared with standardized product databases, this content can provide more specific details: whether a product is convenient to use, whether a store's service has changed, and whether a piece of clothing suits a certain body type.

It is undeniable that this information is very valuable for consumption decisions. But the problem is that a note records one person's experience at a certain time and under certain conditions. The note is real, but it may not be universal; it can provide a reference, but it may not necessarily be directly transformed into an answer suitable for another person.

For example, a coffee machine may receive a large number of positive reviews because its extraction performance is good. But for someone who only drinks coffee two or three times a week and does not like cleaning the machine, this product may not be worth buying.

The value of consumption experience lies not only in how many people say it is good, but also in whether AI can know who is saying it, why they are saying it, and whom this experience applies to.

For Xiaohongshu, the importance of this point is self-evident.

In September this year, Xiaohongshu disclosed that from June to August, the platform cleaned up and handled nearly 120,000 fake notes produced and published by AI. Some accounts used AI-generated images to fabricate experiences such as store visits, accommodation, medical treatment, and beauty treatments, inventing firsthand experiences in an attempt to mislead users' decisions.

In the past, such fake notes mainly affected ordinary users' search and browsing experience. If users saw a fabricated store-visit note, they might make a wasted trip because of it; if they saw a forged product experience, they might make a wrong purchase decision.

And when AI begins to directly read, organize, and summarize this content, the problem becomes even more complicated.

When ordinary users browse notes, they can still judge whether the content is credible by themselves through the author's identity, image details, the comment section, and differences among various notes; if AI directly compresses a large amount of information into a few conclusions, the differences in sources and disagreements of opinion in the original content may be weakened.

Even more, when AI-generated content begins to enter the community and is then used by other AI as a source of experience, low-quality information may form a loop between content production and information retrieval.

For "Mishu," what needs to be done is not limited to helping users provide opinions. It needs to identify information sources and learn to distinguish commercial cooperation content, ordinary user experience, and AI-assisted generated content; when a product has a large number of positive or negative reviews, it should be able to retain different judgments; and it should distinguish the timeliness of information, especially for needs such as dining, travel, and offline services, where past experience may not represent the current situation.

These capabilities may sound less eye-catching than real-time multimodality and long-term memory, but they directly determine whether users are willing to trust AI's suggestions.

In addition, "Mishu" has an even more important task.

If its goal is only to help users find popular products and popular restaurants more quickly, then the higher the recommendation efficiency, the higher the seeding efficiency may also be. But if it wants to become a true personal assistant, it needs to consider whether users really need this product and whether there are more suitable alternatives for them.

Seeding solves why users want to buy, while a personal assistant often needs to answer whether users should actually buy.

A product is very popular, and AI can tell users why it is popular; but when users have a limited budget, use it infrequently, or already own similar products, AI should also be able to give advice to postpone the purchase.

However, this point, which is valuable to users, also means that for the platform, the meaning of consumption conversion needs to be reunderstood.

In the past, Xiaohongshu's commercial value largely came from helping brands find suitable audiences, stimulating interest through content, and then converting interest into purchasing behavior. Now, Mishu needs to gradually begin trying to stand on the consumer's side, helping users screen products, compare experiences, and even identify things that are not worth buying.

A valuable consumption Agent cannot only count how many purchases it has facilitated; it should also pay attention to how many inappropriate consumption decisions it has helped users avoid.

If "Mishu" is viewed together with Xiaohongshu's commercialization moves this year, its significance is actually more than that of an AI product.

On August 7, Xiaohongshu released the NEXT AI commercial product solution, extending AI capabilities to multiple links in merchants' operations, from user insights and content production to advertising placement and transaction conversion.

Simply put, NEXT helps merchants better understand consumers, produce marketing content, and run their businesses; Mishu hopes to help consumers understand needs, screen information, and make choices.

This is quite interesting: merchants are beginning to use AI to improve the efficiency of influencing consumers, and consumers are also beginning to use AI to improve the efficiency of judging products.

Previously, Xiaohongshu mainly connected buyers and sellers through recommendation algorithms: merchants studied the platform's traffic rules and optimized products and content; consumers browsed notes, compared reviews, and then decided whether to buy.

Now, AI may gradually change this chain.

Brands can use AI to analyze user preferences and determine what selling points are more likely to attract attention; consumers can also use AI to organize reviews, compare products, and even directly ask whether a certain product suits them.

Both sides have gained stronger information-processing capabilities, and the platform in the middle can create synergy, helping merchants operate while also helping users make decisions.

However, a clear boundary line needs to be established between these two sets of AI.

Consider a scenario where a brand uses NEXT to analyze a product's core consumer base and continuously produces content targeting these groups. A user then asks Mishu whether the product is worth buying. How should the AI handle this content?

If it makes recommendations based solely on the number of posts, engagement metrics, and frequently cited selling points, then while merchant-facing AI improves content production efficiency, it may simultaneously compromise consumer-facing AI's ability to judge information objectively.

The platform therefore needs to give consumers transparency: why the AI recommends a given product, what content it drew upon, whether commercial partnership information is involved, and whether dissenting user experiences exist.

After all, just as merchants hope AI will help them sell better, consumers also hope AI can independently help them make informed judgments. Otherwise, AI is merely repackaging past information-flow ads in a different form—and what users see is still whatever the platform wants them to buy.

Moreover, creators on the Xiaohongshu platform may not particularly welcome such changes.

Previously, users came to Xiaohongshu and gathered information through individual posts. Creators earned views, saves, and interactions, and the platform sustained community engagement through content distribution. But when Mishu can directly read a large volume of posts and synthesize them into a concise answer, users may no longer need to browse through them one by one.

For consumers, this represents an efficiency gain; for creators, however, the way their content is used has fundamentally changed.

A post may be repeatedly invoked by AI without necessarily bringing the author corresponding clicks and interactions. If users only remember the answer AI provided but never learn whose experience it came from, creators' motivation to keep sharing may be undermined.

Of course, if AI can accurately label information sources, guide users back to the original posts, and give more exposure to content featuring unique experiences, authentic details, and professional judgment, Mishu may also open up new opportunities for high-quality creators.

The key lies in how the platform confronts and manages AI's use and distribution of community content.

From this perspective, the birth of Mishu not only adds an AI entry point to Xiaohongshu but—while transforming how users consume information—may also reshape how merchants operate their content and even affect the distribution of value among creators within the community.

Looking back at the evolution of the internet—from portal websites and search engines to super apps—every shift has brought about a realignment of how industry benefits are distributed.

But entry points have never emerged out of thin air. Search engines established their value through information indexing, e-commerce platforms built barriers through product and transaction systems, and short-video platforms reshaped user attention through content distribution and algorithmic recommendations.

In the Agent era, the original division of labor may once again be disrupted. In the past, users had to actively seek out platforms, and platforms decided what to show them; in the future, Agents can find information, compare products, invoke services, and even decide on users' behalf which platforms they ultimately interact with.

Even if a platform possesses valuable content and services, it may not necessarily retain control over the user entry point.

The emergence of Mishu represents Xiaohongshu's attempt to extend its own content advantages to the user service entry point. After all, content being in demand does not mean that the platform providing that content will necessarily be in demand.

*The cover image and illustrations in this article are sourced from the internet.