Author | Wen Yehao, Wu Xianzhi

Editor | Hao Xin

The rollout of consumer-facing AI agents is like a wedding: hardware makers have their own worries, and software providers have their own plans. Earlier this year, Qianwen and Doubao showcased two very different strategies.

Qianwen’s approach centers on its own app ecosystem. Facing challenges in external partnerships, it turned inward. It created a standalone Qianwen app on the front end and integrated Alibaba’s service ecosystem on the back end. This way, it avoided direct competition with external manufacturers while strengthening its existing ecosystem to deliver services. The newly launched 'Qianwen Services' highlights Alibaba’s vision for user-facing agents, offering features like food delivery, online shopping, ride-hailing, hotel booking, and ticket purchasing.

Doubao, in contrast, took a different route by teaming up with hardware makers to build an operating system-level agent. It skipped the bride’s close friends and aimed straight for the in-laws. Using GUI technologies like screen reading and simulated clicks, it 'understands' user actions to enable higher-permission agent operations. However, this approach has sparked concerns about bypassing third-party software providers, leading to significant pushback.

When it comes to hardware makers and software providers, Qianwen and Doubao each have their biases. WeChat, however, has chosen a middle path. Its A2A (Agent-to-Agent) solution eases hardware makers’ worries about operating permissions, reassuring the in-laws. Meanwhile, the AI Mini Program offers interfaces for software providers, giving the bride autonomy after marriage.

According to the Science and Technology Innovation Board Daily, WeChat has started working with major domestic phone makers to introduce A2A assistant features.

The functionality will revolve around the A2A collaboration mechanism. The phone maker’s AI assistant sends instructions to WeChat, which executes them and returns the results. Currently, some Honor phones are the first to support this feature. From practical experience, users can activate YOYO via voice and issue WeChat-related commands, such as sending voice messages or making WeChat calls.

Additionally, WeChat will roll out an AI agent product, allowing users to access all services within the WeChat ecosystem through natural language. Given WeChat’s massive user base and complex ecosystem, the WeChat Agent may not be fully launched but will instead fit into a standalone AI Mini Program ecosystem.

WeChat: Not Just a Traffic Hub

'I believe many programmers dream not just of writing a program but of creating one that can run other programs. That’s pretty cool.'

On December 28, 2016, Zhang Xiaolong stated clearly at the WeChat Open Class that WeChat Mini Programs were designed as 'programs that run programs,' though he also mentioned he didn’t believe they could develop into an operating system. Ten years later, the now highly mature WeChat Mini Program commercial ecosystem faces reconstruction and challenges from AI.

This January, WeChat Open Class Pro revealed data showing that the WeChat ecosystem has 1.07 billion monthly active users, with Mini Programs covering 108 niche industries. According to a QuestMobile report, among mainstream Mini Program ecosystems, life services are the most active sector across WeChat, Alipay, and Baidu.

The vast scale and complex commercial ecosystem are both the most crucial support and the heaviest burden for WeChat’s AI transformation.

WeChat serves about 1.4 billion users, covering core scenarios like interpersonal communication, payment transactions, government services, medical services, and travel ticket bookings. Any feature embedded in WeChat carries additional risks, as misjudged instructions, incorrect bill payments, or misrouted service ports could be amplified due to the 1.4 billion user base.

Earlier this year, Tencent filed a DMCA complaint with GitHub, demanding the removal of open-source projects related to exporting or analyzing WeChat chat records. Tencent responded that these projects could reverse-engineer the client, crack local database keys, bypass encryption measures, and threaten user privacy.

Besides user scale, WeChat’s introduction of AI Mini Programs is, to some extent, an attempt to disrupt the existing business model of Mini Programs as little as possible.

WeChat Mini Programs offer two types of commercial distribution: service distribution, where third-party providers like Meituan, JD.com, and Ctrip convert traffic into transactions through WeChat Mini Programs; and advertising distribution, where games and short dramas primarily monetize through ads.

It’s clear that the underlying mechanism of Mini Programs aligns well with the invocation logic of AI Agents. Mini Programs are standardized, schedulable service units, allowing WeChat to avoid the need for system-level permissions like Doubao to understand user needs or being confined to Qianwen’s proprietary ecosystem. Theoretically, any long-tail service within the WeChat ecosystem can be invoked through a single voice interaction.

It’s understood that WeChat’s AI agent project has entered the prototype testing phase. Users can activate it by swiping right on the main WeChat interface and complete interactions through natural language. Currently, it offers capabilities like intelligent retrieval, price comparison, filtering, and ultimately completing transactions—all without manual intervention or redirects.

In the era of Mini Programs, WeChat acted as a pure connection platform. Users had to actively search and click, with WeChat merely setting the rules and distributing traffic for the ecosystem. Relying solely on distribution, the WeChat Mini Program ecosystem has nurtured giants like Tongcheng and Luckin Coffee.

After the deployment of agents, WeChat cannot remain limited to connections but must provide a comprehensive set of actions, including user intent understanding, service invocation, task decomposition and orchestration, and secure environments. WeChat will be closer to user intent and, consequently, closer to the operating system itself.

After receiving user instructions, the large model completes intent recognition, extracts key elements from the instructions, and parses tasks. AI Mini Programs can be viewed as toolkits for the execution phase. In addition to providing corresponding services, WeChat also needs to offer other capabilities to ensure instant responsiveness, such as service discovery and matching, and standardization of heterogeneous interfaces.

The core capability provided by the WeChat Agent is task decomposition and orchestration. For complex instructions, the Agent needs to break them down into multiple executable steps, match them with corresponding services one by one, and intelligently adjust subsequent plans based on feedback.

Following this logic, AI Mini Programs may replace the service distribution function of traditional Mini Programs, potentially without fundamentally altering the commercial model of advertising distribution.

Mobile Phone Makers’ Fight to Protect Their Turf

For phone makers, the A2A approach may lack bold imagination but represents a route more aligned with their interests.

Over the past few years, the narrative around AI phones has seemed lively, but attempts such as memory-connecting features and agents have not broken free from the usual upgrade paths of consumer electronics, resembling more incremental innovations than disruptive ones.

In contrast, Doubao phones are far more radical. Despite not fully realizing their potential amid pushback from other players, the waves they have created are sufficient to alert the mobile phone industry.

For a long time, while competition among phone makers has been fierce, as a community of interests, they have also shared an unspoken understanding—competing fiercely on the front stage while remaining united against external threats. Regardless of how market shares change, it is preferable to keep the same players at the table, and they rarely kick each other when down.

Whether it was the early 'Zhonghua Kualian' (China Mobile, Huawei, Coolpad, Lenovo) or later alliances like the 'Hardcore Alliance' and 'Mutual Transmission Alliance,' there has always been a sense of solidarity against common threats.

The arrival of Doubao phones signifies new possibilities for software-hardware coupling, to some extent breaking this unspoken understanding. For phone makers struggling to tell new stories, joining Doubao phones may seem attractive narratively, but behind the allure lies a deeper fissure.

In the short term, collaborating with Doubao can quickly enhance AI capabilities and attract significant attention. However, in the long run, it means ceding significant control over the critical entry point.

Mainstream phone makers are clearly reluctant to accept such a demotion.

If core AI capabilities come from external sources, manufacturers may ultimately become mere hardware 'skeletons,' like ASUS or HP in the PC era, completely losing their voice and merely manufacturing machines. However, some players may not be able to resist the temptation.

Before WeChat’s arrival, mainstream phone makers had already fallen into a prisoner’s dilemma.

If one manufacturer takes the lead in joining the Doubao camp, it may accumulate significant momentum in the short term. However, this advantage will not be exclusive. Once one succeeds, others will quickly follow, ultimately leading to a collective surrender of the steering wheel as everyone integrates with Doubao, erasing differentiation. In contrast, collectively resisting the temptation might be the optimal solution.

However, based on past strategies, such collective rationality often does not last long among phone makers. Moreover, with the industry currently stagnating, some players are already in dire straits, waiting to be 'saved.'

For example, Honor has seemed somewhat 'lost' in the mobile phone arena over the past two years, lacking clear anchors and experiencing declining market share, compounded by the pressure of going public. Naturally, they are anxious.

In April this year, media reports indicated that Honor had been in talks with ByteDance about jointly developing a new generation of Doubao phones. Although Honor denied this, the news inevitably stirred suspicion among manufacturers.

Just as the industry was tensing up, WeChat’s arrival seemed to offer phone makers a better choice.

On the one hand, compared to ByteDance, which has long coveted hardware entry points, Tencent is a more moderate partner, lacking the ambition to 'seize the soul' and allowing manufacturers to retain their AI capabilities.

On the other hand, as mentioned earlier, WeChat is already the closest thing to infrastructure in China’s mobile internet, covering contacts, group chats, payments, Mini Programs, local life services, and even substantial offline commercial scenarios. In a sense, integrating with WeChat is akin to seamlessly connecting to most of the traditional internet’s main arteries without significant disruption.

This represents a more balanced alignment of interests—phone makers safeguard their entry points, WeChat extracts value, and users gain an improved experience. At least on the surface, all parties have a way to save face. Consequently, compared to their cautious approach toward Doubao phones, manufacturers like Huawei, Honor, Xiaomi, OPPO, and Vivo have been much more rapid in responding to WeChat’s olive branch.

The AI 'Wedding': Renegotiating the Terms

Whether it’s WeChat’s A2A or Doubao phones, it is unlikely to be a single-choice question.

Even if manufacturers embrace WeChat enthusiastically at this stage, it does not rule out the possibility of players switching to ByteDance’s camp in the future.

This means the value of this collaboration may far exceed the collaboration itself. For phone makers, a controlled, clearly defined collaboration could alter their approach to partnerships, serving as a bargaining chip in future negotiations with other AI giants.

After all, if even WeChat, which once wielded significant influence and even demanded that automakers reserve a hardware button for 'boarding,' can collaborate within the A2A framework, why can’t other players?

In other words, on the surface, this collaboration is about embracing WeChat, but in reality, it is about using WeChat to stabilize themselves. Phone makers previously had no choice, but now they have more options. Consequently, the 'terms' for mainstream phone makers allying with Doubao may need to be renegotiated.

On the other hand, WeChat’s arrival raises another question: if phone makers are unwilling to cede AI dominance to Doubao, might ByteDance simply build a true Doubao phone itself?

As hardware becomes a new distribution entry point and AI a new interaction layer, model manufacturers naturally feel the urge to extend into terminals.

From PICO in the metaverse era to AI headphones, Doubao phones, Saido cars, and the yet-to-be-released AI glasses, ByteDance has consistently explored hardware entry points without entirely giving up. At present, the smartphone supply chain is highly mature, and hardware manufacturing itself is no longer an insurmountable barrier. If ByteDance seriously considers entering the hardware arena at some point, it would not be surprising.

Ultimately, the present collaboration involving WeChat serves as nothing more than the opening chapter of this grand transformation. In the foreseeable future, the most pivotal narrative within the AI terminal industry will revolve around the alliances and rivalries that unfold among major industry players and hardware manufacturers.