---
title: "ByteDance's Doubao Wanted to Click for You. 150+ Apps Said No."
date: 2026-11-19
category: AI in Action
site: NeuroAI
canonical: https://neuroai.site/a/na-app-doubao-phone-agent-app-blocklist
language: en
---

# ByteDance's Doubao Wanted to Click for You. 150+ Apps Said No.

> When a 30-day comment window closed in October 2026, more than 150 apps formally refused to let ByteDance's Doubao phone assistant operate their screens — exposing a fight over who really controls a user's intent.

You tell your phone you want a train ticket. It opens a travel app, compares times, fills in your details, and stops at the payment screen for you to confirm. That is the promise of the personal agent (个人智能体): a piece of software that does the tapping so you do not have to.

In China this autumn, that promise ran into a wall made of other apps.

On October 14, 2026, a 30-day public comment period ended for a protocol filed by ByteDance's Doubao phone assistant (手机助手). What followed, in the Chinese coverage and in lists circulating among users, was a tally of more than 150 apps — 155 by one count — that had formally declined to let the assistant drive their interfaces. WeChat, QQ, Taobao, Alipay and many banking and carrier apps were on it.

The apps still install, log in and work perfectly well when a human touches them. What they refused is narrower and more consequential: letting an artificial intelligence read their screens and inject clicks and swipes on the user's behalf.

## Key takeaways

- The refusal is about **screen automation**, not access. Apps on the list remain usable; only Doubao's ability to operate them through the graphical interface was declined.

- More than 150 apps said no, including 66 from the Tencent ecosystem and 23 from the Alibaba ecosystem, according to reports by China Newsweek and The Paper.

- The flashpoint is a protocol called SAEP, which ended its comment window on October 14, 2026, after the consumer version of the assistant shipped in mid-September.

- The deeper fight is commercial: an agent that chooses which app to open and which product to show decides who gets the click, the ad view and the payment.

## What SAEP actually proposed

SAEP stands for Screen Automation Execution Protocol — in Chinese, 屏幕自动化操作声明协议. It is ByteDance's attempt to set rules for how an agent may act inside someone else's app.

The framework is more careful than the phrase "screen clicking" suggests. China Newsweek describes identity checks for the agent, verification of package names and signatures, policies that can be set at the app level or the page level, a requirement that the user take over for actions such as payment, and an audit trail of what happened. In other words, the assistant would show who it was, ask before doing anything sensitive, and leave a record.

That is not the same as asking for a blank cheque. But from an app maker's point of view, the core problem remains: the agent sees the screen the way a person does, and there is no pre-agreed list of what it may read or touch. A supermarket cannot decide in advance which shelf the shopper's proxy will look at.

## Why the apps said no

Three concerns surface repeatedly in the Chinese coverage, and none of them is exotic.

First is security and identity. An app cannot easily tell whether a click came from the phone's owner, from an authorised agent, or from something hostile controlling the handset. Yu Bo, chief engineer at the Zhongguancun Super-Interconnection Alliance, told China Newsweek that if something goes wrong, pinning down responsibility is extremely difficult.

Second is commercial control. He Yuan, executive director of the data-law research centre at Shanghai Jiao Tong University, argues that the real shift is not how an ad is displayed but who discovers the user's intent and allocates the transaction. An agent that understands "I want a quiet washing machine for a narrow balcony" and picks the result quietly demotes the platform from a storefront to a back-end supplier.

Third is liability. Payment, messaging and banking apps carry rules about authentication and fund safety. Letting a third party automate a transfer or a message raises questions the existing rulebook was not written for.

Importantly, the refusals are not a blanket rejection of cooperation. Several of the same apps have signed targeted deals with other agents that work through agreed interfaces. He Yuan notes that accepting an API while rejecting broad screen simulation can be a defensible technical position, because an interface lets the app define its own boundaries.

## Two roads, and a number to watch

The industry is split between two designs. One is the "eyes and fingers" route: read the screen, mimic a human tap. It can theoretically reach almost any app without prior permission, which is why handset makers like it. The other is the "agreed door" route, built on protocols such as MCP and A2A, where each app decides what the agent may do.

The numbers explain the urgency. IDC estimates that 2026 shipments of a new generation of AI phones in China will reach 147 million units, up 31.6 per cent year on year from about 112 million units in 2025, with penetration above 53 per cent. QuestMobile put AI-native app monthly active users at 499 million as of May 2026, up 85.4 per cent year on year, with Doubao, Qianwen and DeepSeek forming one of the few groups above 100 million users.

A phone that can act is arriving faster than the rules for what it may act on.

## Honest limitations

This article draws on reporting by China Newsweek (October 21) and The Paper, plus a Reuters summary carried by Taiwan's Economic Daily. I have not seen the full SAEP document or the complete refusal list, so the ecosystem breakdown — 66 Tencent apps, 23 Alibaba apps and so on — is as reported, not independently counted. The totals differ slightly between outlets (150-plus versus 155), and I have kept the smaller phrasing in the headline.

The IDC and QuestMobile figures are third-party estimates quoted in Chinese media, not company filings, and shipment forecasts can move. I have not interviewed ByteDance or any of the refusing apps; the explanations of motive come from named academics quoted in the original reporting, and should be read as reasoned interpretation rather than confirmed internal thinking.

The comparison with other agents is drawn from the same reports. Chinese concepts and protocol names are transliterated where the original wording matters, and English renderings vary.

## What readers can do now

- **Treat "AI can do it for you" as a permission question, not a model question.** Before trusting an agent with a task, check which apps it actually controls and which it merely opens.

- **Keep payment and messaging under manual control.** Even well-designed protocols in this story route high-risk actions back to the user; make that your default across all agent tools.

- **Watch the standards, not the launch demos.** The outcome here will be decided by published, verifiable interface rules — MCP, A2A and whatever replaces them — rather than by how slick the next phone commercial looks.

The phone in the demo can book the ticket. The interesting question is who gets to decide which ticket it sees first. That argument is only beginning.

**Sources:** China Newsweek (中国新闻周刊), Oct 21, 2026; The Paper (澎湃新闻), Oct 14, 2026; Reuters summary via Economic Daily (经济日报), Oct 26, 2026.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-doubao-phone-agent-app-blocklist
Free to quote with attribution and a link to the original.
