The Agent Wars: ByteDance, Alibaba, Tencent and Smartphone Makers' Next Zero-Sum Game
Competition among big tech agents has shifted from model capabilities to tool calling, long-chain fulfillment and system permissions. As of September 2026, domestic AI-native apps reached 546 million monthly active users, with Doubao at 419 million, Qwen at 158 million and DeepSeek at 123 million; Doubao's smartphone assistant screen-automation operation refusal list already includes 155 apps. Ant Lingguang 2.0, one week after launch, managed over 1.16 million affairs and executed over 3.6 million tasks, and has integrated more than 40 external services including DingTalk and Feishu.
Body · AI translation
Around the Start of Winter, the clashes among major tech companies in the agent arena suddenly heated up.
On November 6, Honor CEO Li Jian announced in Guangzhou that Honor will join forces with China Mobile and other partners to build an agent service network, and plans to try launching AI Token packages on the Magic9 series phones.
A day earlier, Ant Group released the report card for Lingguang 2.0 one week after its launch. According to data disclosed by Ant Group, Lingguang has already helped users manage more than 1.16 million items and cumulatively executed more than 3.6 million tasks, and has connected to more than 40 external services including DingTalk, Feishu, and Hello.
At the end of 2025, a Nubia engineering prototype equipped with the Doubao phone assistant made its debut; Doubao can recognize screen content and attempt to operate the phone on the user's behalf. In October this year, the rejection list for screen automation operations published by the Doubao phone assistant already included 155 apps, including commonly used software such as WeChat, Taobao, Meituan, and Xiaohongshu. Users can use them normally themselves, but Doubao cannot freely operate them on the user's behalf.
As a result, the arms race among large models has quietly shifted battlefields. As answering everyday questions no longer creates a gap, the industry's focus has rapidly shifted to tool invocation, long-chain fulfillment, and system permissions.
QuestMobile data shows that as of September 2026, the monthly active users of domestic AI-native applications had climbed to 546 million, and users of terminal system-level AI had crossed the threshold of 780 million.
The user base is large enough, but the rules of the game are being rewritten. Within the existing landscape, China's internet giants each ruled their own domains through search, social networking, e-commerce, and algorithmic recommendations, while at the same time facing a new division of interests.
The Battle for the Entry Point
Over the past two decades, the map of China's internet, in the final analysis, has been about dividing up the power to distribute demand.
The boundaries between the domains each company ruled were very clear. Baidu guarded the search box, Alibaba controlled commodity transactions, Tencent locked down interpersonal networks, and ByteDance used algorithmic information feeds to continuously carve up users' time. Everyone built their own high walls, collecting taxes steadily within their own territories, and coexisting peacefully for many years.
The path to finding information and getting things done has suddenly become much shorter. Users no longer need to flip through web pages one by one in a search engine, but are accustomed to simply throwing a sentence to Doubao, Qwen, or DeepSeek. Asking the model when there is something to do is becoming muscle memory.
QuestMobile's data is very intuitive. In September this year, Doubao's monthly active users crossed the 419 million mark, Qwen reached 158 million, and DeepSeek also climbed to 123 million. Looking solely at user scale, Doubao has indeed pulled ahead among domestic AI-native products.
But scale does not directly equal victory. The larger the user base, the more obvious ByteDance's sense of anxiety becomes.
A person opens Doubao dozens of times a day to chat, writes a few paragraphs of material, and edits two emails. It seems highly sticky, but the liveliness in the dialog box will not conjure cash flow out of thin air. A user asks Doubao to painstakingly plan a travel itinerary in clear detail, but in the end, when choosing a hotel and buying a plane ticket, they still have to jump out of the interface, and the orders and handling fees ultimately land in the pockets of OTA platforms such as Ctrip and Fliggy.
This is exactly where the real calculation behind the Doubao phone assistant lies. ByteDance does not want to be just a chatty AI that answers questions and resolves doubts; it wants the agent to reach directly into the operating system. By recognizing phone screens and simulating human clicks, the operations users originally had to perform back and forth among Meituan, Ctrip, and Taobao to compare prices and repeatedly jump between pages are all handed over to Doubao to handle. To put it plainly, ByteDance is urgently attempting a boundary-crossing interception, trying to turn its own AI into the master remote control for all apps, and to stake out a position in advance at the very front end where users actually make payment decisions.
Alibaba's trump card and mindset are completely different.
Taobao, Tmall, Amap, Fliggy, and Ele.me form a complete commercial fulfillment network covering physical e-commerce and local life. For Qwen, even if it plays only an inconspicuous wake-up entry point on the consumer side, as long as users initiate consumption demand, orders will most likely still land steadily in Alibaba's own commercial flywheel. Alibaba's real challenge is not finding services externally, but how to coordinate old and new interests. Taobao has long earned rich profits from merchants' bid-based search, recommendation exposure, and advertising slots, while Qwen is committed to helping users compress the time spent comparing and choosing. Once users fully rely on agents to make decisions, the original business logic of making money from impressions will have to face deep restructuring.
Tencent's choice has another kind of composure and restraint.
What WeChat has accumulated is not only a massive social relationship chain, but also a decentralized service foundation built by mini-programs. People are already used to chatting, getting things done, and making payments in WeChat. For Tencent, there is no need to aggressively build a cross-software plug-in like ByteDance and probe other companies' interfaces everywhere. As long as agents are embedded silently into chat boxes, Moments, and enterprise collaboration systems, and its own moat is defended, it will be very difficult for others to break in. The only thing Tencent needs to carefully weigh is exactly how much operational authority to release to agents, and which core links must be firmly held in WeChat's palm.
Of course, there is another force in this card game that cannot be ignored, namely the phone manufacturers that hold hardware terminals. Compared with third-party software, phone manufacturers naturally occupy dominance at the bottom layer of the system. Wake-free operation, cross-software scheduling, and all-weather hardware-level perception—Honor teaming up with China Mobile to launch an agent service network is essentially an attempt to use the first touchpoint of hardware devices to retain users' AI entry point on the system desktop.
The positions several major companies occupy at the entrance are entirely determined by the assets each has accumulated over the years. ByteDance is eager to expand outward in search of real-world application scenarios, Alibaba is busy using AI to rebuild its own transaction fulfillment, Tencent is carefully protecting the order of its existing ecosystem, and phone manufacturers are striving to hold the physical choke point of devices.
The battle over the entrance has thus split into two starkly different real-world paths. One is independent AI applications relying on conversational interaction, striving to become all-powerful super hubs; the other is super apps and mobile operating systems relying on their existing turf, quietly sinking AI capabilities down into infrastructure.
In the short term, user scale on the software side more easily builds momentum on the rankings, but in the long term, whoever can make users feel comfortable handing over deep permissions, and whoever can be awakened first in daily life, will have the confidence to set the rules for the next generation.
Ecosystem battle
As for the "list of 155 rejections" that Doubao ran into, it lays bare the cruel reality of agents' deployment: no matter how smart the model is, without platforms opening up data and interfaces, it can hardly move a single step.
Booking a flight ticket requires checking flights, verifying prices, filling in identity information, and finally completing payment. The whole process involves multiple systems, and if any one link cannot obtain the necessary data or operational permissions, the task may be interrupted. Even if AI can accurately understand the user's request, it may not have the ability to complete the actual operation.
The problem is that most of the services agents need to call are controlled by other internet companies.
For a long time, the business model of the mobile internet has been built on the cornerstone of "humans using apps manually": people open Taobao to find products, browse Meituan to pick merchants, and choose hotels on Ctrip. Platforms influence user choices through recommendations and advertising, and merchants, in order to gain traffic and orders, are also willing to pay the corresponding fees.
Agents may change the original transaction process.
For example, when a user is preparing to buy a computer and tells AI the budget and configuration requirements, AI can query prices, after-sales service, and delivery conditions across different platforms to help the user complete the comparison. Users save the trouble of repeatedly browsing product pages, and the opportunities platforms originally used to display ads and recommend products may also decrease accordingly.
For ByteDance, if Doubao can complete more tasks on users' behalf, it has the opportunity to reach more consumer demand. For platforms such as Taobao and Meituan, letting other companies' AI directly operate their services also means the original user entrance and product recommendation mechanisms may be affected.
The interests of the two sides are not entirely aligned. Companies providing agents want to connect more services and capture the demands users raise; platforms that own products and transaction systems, on the other hand, must consider how much impact opening their services will have on their original commercial revenue.
Of course, platforms refusing automated operations also has security reasons.
WeChat and Alipay hold large amounts of personal information, involving identity authentication, fund transfers, and account security. Agents that can recognize screens and simulate clicks make things convenient for users, but also increase the risk of erroneous operations and information leaks. A user authorizing AI to handle affairs does not mean third-party software can operate platform accounts without restriction.
Therefore, platform restrictions on automated operations involve not only anti-fraud and account security, but also commercial interests that need to be weighed. It cannot simply be assumed that major companies reject agents in order to protect their own traffic.
This has given rise to two construction paths that will not yield to each other: one is "breaking external walls," represented by Doubao, which tries to formulate operation protocols and force its way through multiple software by simulating humans; the other is the secure closed loop taken by Ant's Lingguang, WeChat mini programs, and Alibaba's Qwen, which relies on official API interfaces and connects cautiously within mutually recognized commercial and security frameworks.
What is more interesting is that major companies generally find themselves caught in a dual role in the competition. They all crave a pass for their own agents into external software, yet they are extremely cautious about opening up their own core assets. It is impossible for WeChat to open its social and transaction interfaces to competitors without limits, and Taobao must also weigh whether external AI acting on users' behalf will eat into its original traffic dividends.
Every company hopes its own agent can connect to more software in the external ecosystem, yet when it comes to opening its own core business, it needs to weigh things repeatedly. At this point, agents have completely broken away from algorithmic self-indulgence and stepped right into the highly sensitive interest zone among major companies.
How far interfaces should be opened, how user privacy should be verified across endpoints, who will bear the losses when there are bad debts or mistaken operations, and how revenue-sharing ratios should be split—each item must be repeatedly haggled over at the negotiating table. Even if model vendors push reasoning capabilities to the extreme, they still cannot pry open the moats others have built over more than a decade. To put it another way, even if users trust a certain agent one hundred percent, if the platform on the other side does not allow it through, the task will still break down halfway.
Even putting that aside, even if all parties really sit down and sign agreements, there is still a computing power bill hanging in the background. Every time an agent completes a long-chain task for a user, behind it is real computing power consumption of real money. The more frequent the calls, the tighter the server costs bite. Being able to run through is one thing; being able to make the numbers work and earn a profit is another life-or-death line.
Business battle
Agents can catch users' demands and can also connect services between platforms. But shifting the gaze from personal phones to enterprise-level collaboration, the battlefield turns from competing for attention to stricter calculations of business efficiency.
QuestMobile data shows that in September this year, China's monthly active users of AI office tools had climbed to 142 million. Tencent brought out WeCom and WorkBuddy, Alibaba relied on DingTalk and Qwen Office, and ByteDance built knowledge bases on Feishu. All three companies are using their original foundations, trying to let agents take over approvals, reimbursements, and cross-department collaboration.
But when it truly lands inside enterprises, reality is far more trivial and stringent than generating a document.
To process an expense reimbursement form, it has to circulate among the employee, the business supervisor, finance, and the cashier; to sign a contract, one must check the client's qualifications, payment terms, and seal-use authority item by item. If an agent wants to get things done on people's behalf, it must not only penetrate the many separate business databases, but also ensure that every step of the operation does not exceed authority or make mistakes.
Bosses do their accounting in a very realistic and naked way. What they want is deterministic results and clear attribution of responsibility. Data cleaning that originally took three people can be handled by one person with AI, and the saved labor costs can be put into the financial statements—that is why enterprises are willing to pay. If it merely generates a document full of empty platitudes that still has to be rewritten manually, the so-called efficiency gain is just an unclear, muddled account.
This is exactly what puts a new cost straitjacket on software vendors.
Traditional SaaS software charges by account and by year, with marginal costs approaching zero. Agents completely overturn this profit model: for every cross-system task processed, the backend must repeatedly call large models and tools, and computing power and tokens are continuously burning real money. The more frequently tasks are executed, the heavier the inference costs borne by service providers. Customers want to pay a fixed fee to buy stable service, but vendors have to secretly calculate whether high-frequency calls will pierce their profit floor. Honor's trial of token packages on mobile phones is precisely an exploration of how to counter this cost black hole.
Looking back at this agent battle spanning both the consumer side and the industry side, the chips of the major companies still come from their original business foundations. ByteDance holds a huge base of AI-native users, Alibaba occupies a deep transaction and fulfillment network, Tencent guards national-level social and collaboration relationships, and phone manufacturers occupy the first touchpoint of system hardware.
However, holding the entrance does not mean being able to complete tasks, and being able to complete tasks does not mean being able to make the numbers work.
The business balance established in the mobile internet era has been broken. A task may have its demand picked up by one company's AI, be served by another company's platform, and finally complete settlement and fulfillment through a third-party channel. Users only care whether the matter is handled properly, while the major companies have to weigh authority, computing power, and revenue sharing at every step.
The endgame of business is extremely plain. When users entrust a matter to artificial intelligence, who exactly gets it done? Who has the final say? And who can steadily earn a profit from this closed loop?
This article comes from the WeChat public account "Xiutai", author: Xiutai, published by 36Kr with authorization.