Informopedia Explains: ByteDance's AI assistant is having the strangest week of its life. On Wednesday, a 30-day countdown expired and its phone assistant began switching on automation inside third-party apps by default. More than 150 of China's biggest apps — the places you chat, shop, pay, and eat — said no.
The assistant is Doubao, the AI app from ByteDance — the Beijing-headquartered tech giant behind TikTok, and one of China's biggest AI builders. Its "Phone Assistant" is the ambitious one: an agent that doesn't just answer questions but does things — books, orders, pays, posts. To do that, it needs to reach inside other companies' apps. This week, those companies drew a line.
What just happened
Doubao published a new rulebook in September called the Screen Automation Execution Protocol, or SAEP. It governs the way the assistant operates other apps: reading their screens and simulating taps and typing. Then ByteDance did something unusual — it flipped the default. After a 30-day public notice period, which ended on October 14, Doubao would gradually enable screen automation in third-party apps unless their developers explicitly opted out or set restrictions. Old rule: the assistant waited to be invited in. New rule: silence means yes.
The opt-out list is the story. National Business Daily counted 151 apps that refused access; other tallies put it at 153. Either way, almost the entire Tencent and Alibaba ecosystems are on the list. The names read like the Chinese mobile internet itself: Tencent's WeChat and QQ; Alibaba's Taobao, Alipay, and Ele.me; plus Meituan, Pinduoduo, Kuaishou, and Xiaohongshu.
And then there's the delicious detail: the refusal list includes Qwen — Alibaba's own AI assistant — plus DeepSeek and Kimi, two of China's leading AI labs. The AI assistants are refusing each other access now.
And one thing to be clear about: you can still install, open, and use every one of these apps yourself. The refusal only stops Doubao from reading and operating their screens. Your WeChat works fine. It's the robot that got locked out.
How an AI agent actually runs your errands
An assistant that only answers questions needs nothing but your words. An assistant that does things needs hands. There are two ways to give it hands.
The first is the front door: official APIs. The app builds a clean, sanctioned hook — "yes, assistant, you may check my order status this way" — and the agent uses it. Safe, limited, controlled.
The second is the wall-climb: screen automation. The agent looks at the phone screen the way you do — reads the pixels, recognizes the buttons, the menus, the payment fields — and simulates taps and typing to get the job done. That's what SAEP governs. It's powerful precisely because it works on apps that never built a front door. It's also technically indistinguishable from what malware does: watch your screen, touch your screen. That duality is the whole problem.
Why the platforms can say no
Technically, it's simple: it's their software and their servers, and an unauthorized agent looks exactly like a bot.
This isn't the first round: when the first Doubao phone went on sale late last year, WeChat flagged users for abnormal logins and Alibaba's apps threw up verification checks — the agent's activity tripping standard security controls. ByteDance publicly pulled back, suspending gaming, banking, and payment functions.
Legally, the platforms stand on familiar ground: their apps' terms of service, Android's permission model, and — crucially — Doubao's own protocol. SAEP was written so that developers can refuse, and ByteDance promised to stop the relevant operations when they do. The veto is a feature of the system, not a bug. That's why the refusal list works at all.
And the platforms are not saying no to everyone. They're saying yes on their own terms. Alipay lets outside assistants pass service requests to its own agent, Abao — come through our agent, not around it. WeChat offers partner phone assistants a narrower route limited to messages and calls. The message is consistent: assistants are welcome as guests, not as landlords.
The real fight: who controls the agent economy
Here's the part neither side will say out loud. The assistant that can act inside every app becomes the new gatekeeper of the phone. Whoever owns the agent owns the user relationship — the orders, the payments, the conversations. WeChat and Alipay spent a decade becoming super-apps, entire economies inside a single icon. Letting ByteDance's assistant see their screens and handle their transactions would mean handing a rival the keys to that economy.
So this is a platform war wearing the clothes of a security debate. The security concerns are real — but notice the pattern: Alibaba blocks Doubao while building its own agent platform, Qwen Intelligence. Tencent blocks Doubao while developing its own assistant, Yuanbao. Every giant wants agents to have hands — just not inside their house.
The global mirror makes the stakes obvious. In the US, Google's Gemini app just gained a "tasks" mode connected to more than 40 apps — access granted by partnership, by invitation, one integration at a time. That's the same war, fought politely. China's version is the same war fought by flipping the default and daring the giants to refuse in public. They did.
And the guardrails race has two fronts now. Anthropic is writing usage rules for how its assistant should be treated by users — governing assistant behavior from the inside. China's platforms are writing rules about what a rival assistant is allowed to touch — governing assistant behavior from the outside. The industry's biggest unanswered question is caught between them: who gets to set the boundaries of a machine that acts on your behalf?
Beijing's shadow: will the regulators step in?
In China, no platform war ends without the referee looking at the field — and Beijing has a record of stepping in.
The antitrust watchdog fined Alibaba a record 18.2 billion yuan in 2021 for locking merchants into exclusivity — when platforms lock each other out, Beijing has shown it will act.
More directly, Beijing has already reshaped this exact category: in April, five agencies issued rules on AI companion services that forced Doubao, Qwen, and Tencent's Yuanbao to shut down user-built companion agents in July. Beijing's stance on AI agents is not theoretical — it has already forced all three giants to rebuild parts of their products.
But no regulator has said a word about the refusal list. The market regulator could treat a coordinated refusal as anti-competitive — or Beijing could decide platform security and data protection justify the walls. The government wants China to lead in AI agents and airtight data security, and those two goals point in opposite directions here.
What to watch: any notice from the internet regulator or the industry ministry about agent interoperability, or any antitrust probe touching the refusal lists. Until then, the walls stand — and the giants' own agents develop behind them.
Why this matters
Step back from the Chinese app names and this is the first great test of the agent era's core question: does your AI assistant work for you, or does it work only where platforms allow?
An assistant that can't open your messaging app, your bank, or your food delivery app is a demo, not a deputy. And the pattern won't stay in China. Every phone maker is racing to build the agent layer — Apple, Google, Samsung — and every one of them will eventually discover what ByteDance just learned: the technology to do your errands is the easy part. The permission to do them is the business. Whoever grants that permission — the app, the platform, the regulator — holds the real power in the age of agents.
What happens next
Three threads to follow. First, the refusal list itself: developers can still submit refusals after the deadline — watch whether it grows, or quietly shrinks as smaller apps decide being inside the assistant beats being outside it.
Second, the moat question. If Tencent's and Alibaba's apps stay closed to Doubao while their own assistants get privileged access, the refusal list stops being a security measure and becomes a competitive moat. That's exactly the pattern antitrust regulators exist to notice.
Third, Beijing. Silence is also an answer: it means the current arrangement — refusal lists, developer vetoes, each giant building its own walled agent garden — is tolerable for now. But China has never tolerated platform walls forever. The last time the giants built walls against each other, the state tore them down. The only question is whether this time the state decides the walls are protecting users — or protecting monopolies.
One Comment