Day 130 of Xiaowei's gray release: WeChat breaks mini programs into parts and hands them to AI

In early October, some users noticed that the mini program homepage in WeChat's "Discover" tab had a new layout: a search box at the top, followed by Recently Used and My Favorites, and below that, separate sections for mini games, short dramas, life services, and AI services. The page now accommodates a search box, category sections, and recommendation slots all at once.

About ten years ago, when Zhang Xiaolong first fully explained mini programs at the WeChat Open Class PRO, he said WeChat would not build a mini program store, nor would it categorize, rank, or recommend mini programs. Over the past decade, as mobile internet entered the stock era and AI layered on top at the application level, both the form and the rationale of distribution have undergone unpredictable changes.

The layout change is only superficial. What WeChat is doing is embedding AI into the entire chain of mini program services—from intent recognition and service matching to order placement and fulfillment. Some of the path choices originally made by users are now partly judged by AI. Counting from the first gray release on June 21, WeChat's native AI assistant "Xiaowei" has been online for 130 days. WeChat officially discloses 16 invocation entry points, while third-party testing has counted 28, covering chats, search, Official Accounts, Channels, and PC.

Xiaowei's entry points spread to 28 locations

In an ordinary user's day, there might be several encounters with Xiaowei: in the morning, tapping the green icon in the upper-left corner of the chat list and asking, "What Official Accounts are worth reading today," and it pushes a few summaries. At noon, long-pressing the search box and saying, "Help me order a lightly iced Americano," and it brings up a mini program and jumps straight to the confirmation page. At night, right-clicking a file in WeChat on PC, and the menu has an extra item: "Ask AI."

These actions happen in different places. Over more than four months, Xiaowei has expanded from a handful of entry points to covering almost all of WeChat's main scenarios. The main entry point in the upper-left corner of the homepage is still in gray release, and the visibility of other locations varies by user. WeChat has not yet announced coverage ratios.

This density is unusual in WeChat's history. In the past, when giving a service an entry point, WeChat was usually very restrained. Payments sit in the Wallet entry, Channels occupies one slot on the Discover page, and for ten years mini programs have mainly relied on developers to drive traffic through their own QR codes. Twenty-eight locations—this degree of openness speaks for itself.

But many entry points do not equal the ability to get things done. Waking Xiaowei takes only two seconds; the hard part is what comes after waking it: understanding what the user says and deciding who to hand the task to. In the normal chain, Xiaowei is responsible for understanding user intent. The model foundation is WeChat team's self-developed WeLM, and for some complex problems it calls DeepSeek. The two work together and handle scheduling. As for fulfillment actions such as taking orders, making products, dispatching vehicles, and booking rooms, the owners of the mini programs must do those.

The fulfillment layer is the only part of this ecosystem that must rely on external parties, because it is done by the mini program owners. But "done by the owners" has a prerequisite: WeChat must first know what a mini program can actually do. WeChat calls this system "Mini Program AI": when a user starts a conversation, it analyzes and selects the appropriate atomic interfaces and atomic components, completes data processing and task execution, and then integrates the results for the user.

In the past, this did not require thought. Users opened a mini program, saw the homepage, menus, and buttons, and completed the operations themselves. But for AI, it needs to know which store, which drink, with what parameters to place an order, and finally what card to present to the user. How this information is obtained is the key to whether Xiaowei can later "get things done."

The "Guidelines for Developers to Connect to the WeChat AI Ecosystem" released on June 8 are precisely about getting developers to fill in this information.

The guidelines offer two paths, and the official wording is that the two modes can be enabled at the same time. Among them, the automatic mode has almost zero threshold: developers check an authorization box, and WeChat reads the source code during review, analyzes the page structure itself, and lets AI operate directly. You can think of it as WeChat sending a clerk into the store to inspect the flow, and then walking through the process on the user's behalf. It is indeed convenient, but the cost is handing over the right to interpret the product: developers cannot intervene in how AI understands its functions, how it presents them to users, or who gets prioritized for the same request.

The development mode requires developers to do the work themselves, breaking the mini program into smaller units—down to separate units for "check store," "check drinks," and "place order." WeChat calls these units atomic interfaces and atomic components, which can be roughly understood as Lego bricks: in the past, when a user tapped "Order Now," behind it was a whole set of page jumps. Now developers must break out each step separately, clearly marking what goes in and what comes out, so AI can assemble them piece by piece. After breaking them out, they also need to provide a manual telling AI how many steps this business has, which block to assemble first at each step, and where the parameters come from. This manual is the SKILL.

The rules for writing this manual are almost harsh, with one core principle: do not make AI guess. For example, parameters cannot be natural language; stores and drinks must be passed as IDs. You cannot expect the model to know which product it is just because the user said "orange C iced Americano." The content returned by the interface to AI also has a fixed format: first explain clearly what happened, then tell it what to do next. You cannot write only an instruction like "show card"; you must write "Order data has been found. Please show the order confirmation card and wait for the user to tap confirm." With only actions and no facts, AI may skip the user confirmation step entirely.

In developers' daily work, the biggest hassle is not writing code, but the implicit rules that used to be "users would just tap it themselves, no need to write it into the code"—now every one of them has to be dug out. For a mini program with simple interfaces, a few lines of description suffice. For a trading system that has been running for years, the experience in veteran employees' heads—"where does this ID come from, and which interface do we call if we can't get it"—all has to be written into a format that AI can understand.

In addition, before starting development, the developer tools must be switched to the Nightly version, the compilation mode changed to "mini program AI compilation," and the base library upgraded to a specified new version. These terms are not normally part of most mini program teams' daily toolchains.

To split or not to split?

Since whether to split is not up to them, all developers can do is weigh the costs and benefits.

On the cost of not splitting, the guidelines put it this way: "Mini programs that have not completed integration will not be recommended or invoked by WeChat AI." The wording is "invoked"; no deadline is set, and no penalty is mentioned. But another sentence in the document also reveals: "After integration, your mini program will have the opportunity to be recommended and invoked by WeChat AI."

But splitting also has costs: the toolchain must be changed, and each atomic interface must be individually encapsulated and individually declared. These deterministic costs are manageable, though; the real problem is that after all this rework is done, at least in the current state, whether your mini program will be invoked is not an extremely "certain" thing.

A development-mode SKILL must pass the platform's evaluation, which is built around four metrics: service delivery, interaction experience, performance quality, and scenario coverage. The first three must score above 80, and scenario coverage above 60, to be eligible for invocation by WeChat AI. The standards are public, but the invocation ranking rules after evaluation have not been disclosed. If the ranking criteria could one day be made public like the evaluation standards, this arithmetic problem would be much easier for developers to solve. All developers can do is write the SKILL well, describe the interfaces clearly, and wait for a result.

For small and medium teams whose main revenue comes from advertising and page conversions, there is another layer to calculate: in automatic mode, AI operates the mini program on the user's behalf, and the user can stay in the conversation the whole time without opening the page even once, so the chance for ad exposure is gone. If AI skips the intermediate pages, conversion attribution also breaks its chain. Some people in developer communities have asked whether AI directly operating pages will break existing tracking and ad display logic. For now, no one has given a clear answer.

So the most likely outcome is a screening: the threshold for integration is so low that everyone will do it, while the threshold for seriously splitting is so high that only a few will split. Top teams with ample capability reserves have already taken the opportunity to break their businesses into parts; teams lacking reserves have a zero-cost fallback—tick an authorization in automatic mode, and that counts as completing integration. Teams in the middle, carefully weighing things, instead have nowhere to apply their efforts. As of now, submission for review is not yet open, and the official documentation explicitly warns not to merge this part of the code into the official version, to avoid affecting normal version releases; the evaluation tool is also being upgraded, and the evaluation SKILL is temporarily unusable.

Who has actually acted first?

However, the fastest movers, the group closest to the inner table, started earlier than the guidelines.

It is understood that Ctrip once confirmed to Caijing that in mid-April this year it had already integrated WeChat AI through development mode, and would subsequently adapt around core scenarios such as hotel booking, flight search, travel vacations, and mobility services. That timing was nearly two months earlier than the June 8 guidelines. By the first publicly disclosed beta list in June, companies including JD.com, Meituan, Didi, Ctrip, Tongcheng Travel, Poizon, and KFC were already on it.

The companies on the list should be going through a channel co-built with WeChat, not public submission for review. This can be inferred backward from the timing: the guidelines were only released on June 8, while Ctrip completed integration in mid-April.

JD.com integrated AI Agents around e-commerce, food delivery, and logistics, and its AI shopping assistant had previously connected with multiple phone manufacturers. Meituan has its self-developed LongCat model (the version opened for testing in June was LongCat-2.0-Preview), with total parameters exceeding one trillion, with training and inference relying entirely on domestic computing clusters; the AI assistant "Xiaomei" based on this model is already making cross-scenario decisions.

Didi integrated its core ride-hailing services, so users can directly summon Express, Premier, and Discount rides through AI interaction within WeChat without jumping between apps. Tongcheng was already a deep partner in the Tencent ecosystem, and after integration users can complete itinerary planning and booking in natural language.

These capabilities were already under construction; integrating with WeChat was just a matter of completing the connection along the way. Whether or not the guidelines were issued, they would have moved in this direction.

Beyond the top players, most small teams choose to turn on automatic mode and wait for the platform to parse them. But this set of rules is new to everyone, and what it takes to be selected is also changing.

In the past, the competition was over SEO, ad spending, and social fission—who could get users to tap in on their own. From now on, it will depend on the quality of interface descriptions and the stability of fulfillment—who can be selected by the machine. The new rules have no column for "historical accumulation." An old mini program that has operated for years with stable users and a new mini program that just launched but has a cleanly written SKILL are equal in the invocation pool. Brand is also weakened: users no longer need to open a particular mini program, and may not even know who is ultimately invoked.

The emergence of a transformation threshold has given rise to intermediary services. On June 10, Weimob became one of the first beta teams to connect to the WeChat AI ecosystem. According to its account, merchants will in the future be able to connect conveniently through Weimob mini programs, with a focus on covering e-commerce, retail, local life, and many other types of merchants. The emergence of intermediaries is itself a signal: transformation really does have a threshold. The fact that merchants are willing to pay to commission transformation shows that they understand its necessity but lack the ability to complete it themselves. "Connection," in this way, has become a business.

The bottleneck is not the model, but the supply of parts

There may be more nominal connections, but the parts that can actually be invoked will not increase as a result, and Xiaowei's actual performance is therefore uneven. Tests by several media outlets have provided samples.

In scenarios with short chains, high frequency, and clear boundaries—such as ordering a cup of coffee or checking an order—the completion rate is relatively high. Compound instructions and tasks involving more than three steps decline noticeably. If this gap holds, the root cause lies more in the supply of parts than in model capability.

This also explains why the three giants are taking three different paths: Tencent bets on the WeChat entry point, ByteDance bets on a standalone app, and Alibaba bets on models and the full stack. Qwen announced in early June that it would fully open to third-party Agents and Skills, and brands such as KFC have already tested Agent services on Qwen.

The three paths actually answer the same question: after a user hands a sentence to AI, who will get it done. Tencent's path has the lowest cost because the fulfillment layer already exists—millions of mini programs running transactions inside WeChat do not need to be rebuilt, only transformed. The bottleneck is not the model or computing power, but how fast the transformation can advance.

Over a longer time horizon, whether this model can work will also depend on two things. One is the mini program MCP: this is the invocation protocol that WeChat AI follows when scheduling mini program capabilities. Through it, AI identifies which atomic interfaces a mini program has available for invocation and what parameters each interface requires, and then dispatches tasks. Its foundation is the open standard MCP proposed by Anthropic at the end of 2024. The agent platforms of Alibaba and ByteDance are both moving toward this standard, but WeChat has built an additional layer on top of it.

According to WeChat's official documentation, the interface forms of mini programs differ greatly from web services, and the standard MCP cannot be directly applied. Therefore, the mini program MCP has been customized according to the characteristics of mini programs: developers must provide a complete SKILL implementation according to the specification before mini program AI can reason correctly and execute the corresponding atomic interfaces. The cost of this customization is that this manual can only be used within WeChat. If other platforms each add their own layer in the future, the same mini program will have to have a manual written for each platform.

Another is compliance. According to industry media reports, the mandatory national standard for agent security, "Basic Security Requirements for Agent Applications," has already been established as a project. Requirements such as behavior traceability, clear authorization, and auditability will sooner or later be counted as development costs. Further ahead, the standard for measuring the value of mini programs may shift from "number of times opened" to "probability of being invoked."

The three months after the development model opens for submission review will be an observation window for making judgments. If by then the share of mini programs using the development model is still less than 10 percent, the conclusion that "only the top players are seriously dismantling" will be settled; if a group of small and medium-sized developers spontaneously rewrite SKILLs and can show attributable incremental invocations, that judgment will need to be revised. What determines the direction may not be WeChat's determination, but developers' actual investment.

Even if the transformation ultimately does not reach scale, this set of specifications will itself remain. How capabilities are described, how parameters are passed, and what format is used for returns—once these conventions have been tested in practice by millions of developers, they will become the industry's de facto standard, and their value does not depend on the success or failure of this transformation.

WeChat has completed its part in these 130 days: the entry points have been laid out and the specifications have been written. Two uncertainties remain: first, the willingness of the owners of millions of mini programs to transform; second, whether those who seriously transform can receive commensurate returns. The answer to the former lies in developer communities, and the answer to the latter lies with WeChat. The speed at which these two variables are realized will determine the upper limit of this ecosystem's scale over the next two years. (This article was first published on TMTPost APP. Author | AGI-Signal, Editor | Jia Yuwei)