TMTPost: Computing Power Is No Longer the Ceiling for Mobile AI Agents—Ecosystem Walls Are

Written by Silicon-Carbon Variables, authored by Hu Yidao

On October 14, the official WeChat account of Doubao Mobile Assistant announced that the 30-day public comment period for the "Screen Automation Execution Protocol" (SAEP) had ended, and the platform would gradually open screen automation operation permissions for third-party applications.

According to the public notice, the latest list of applications refusing operations includes Pinduoduo, Kimi, the Kuaishou series of apps, the Meituan series of apps, Xiaohongshu, DeepSeek, Industrial and Commercial Bank of China, as well as 23 Alibaba and affiliated ecosystem applications including Taobao, Qwen, and Alipay, and 65 Tencent and affiliated ecosystem applications including WeChat, QQ, and Honor of Kings, totaling 151 apps.

According to earlier introductions, SAEP is a set of AI operation rules for third-party apps. Developers can independently declare whether to allow AI assistants to perform screen automation operations on their own applications, and define the boundaries and scope of restrictions. Officials emphasized that after the public comment period ends, app developers can still submit refusal declarations at any time, and upon receipt, Doubao Mobile Assistant will promptly shut down the relevant operations.

The foundation of super apps lies hidden in the traffic closed loop

The computing power anxiety of mobile AI Agents was once the core narrative of the industry for more than two years, but it dissipated faster than many expected.

In just two years, the parameter scale of on-device large models has leaped from billions to tens of billions, and the NPU computing power density of flagship chips has increased substantially; underlying engineering challenges such as memory bandwidth bottlenecks, quantization compression algorithms, and on-device inference frameworks have continued to make progress under joint efforts by chipmakers and phone manufacturers.

Today, many flagship models can already locally support basic AI capabilities such as voice interaction, intent understanding, and image recognition. Multimodal models' perception of phone screens is no longer limited to recognizing button coordinates and has begun to possess the ability to parse shallow page semantics.

Technological iteration is still accelerating, but a core contradiction previously underestimated by the industry for a long time is rapidly surfacing: when AI initially gains the ability to understand screens, simulate clicks, and decompose and execute instructions across applications, what determines the boundaries of mobile AI Agents is no longer the technical ceiling, but the access red lines of the application ecosystem.

Computing power defines the capability ceiling of AI Agents, while ecosystem rules define their actual boundary of activity.

The difficulty of the two is not on the same dimension at all: computing power bottlenecks rely on capital investment and technological iteration and have room for continuous improvement; but the reconstruction of ecosystem order affects the foundation of commercial interests accumulated by the mobile internet over more than a decade.

To understand the motivations behind this group of apps collectively shutting down permissions, it is necessary to return to the core business model that has operated in the mobile internet for more than a decade: the traffic closed loop.

Over the past decade or so, the value chain of mobile internet platforms has always operated around this closed-loop system: platforms use content, social relationships, product supply, and local services as anchors to retain user time and attention within their own app ecosystems, and then monetize value through information feed advertising, search bidding, transaction commissions, membership subscriptions, and other methods.

Every browse, click, price comparison, and order placement by users within an app is a carrier of traffic value; advertisers pay for exposure, merchants pay for traffic, users pay for services, and platforms extract revenue from every layer of value exchange.

The reason this traffic closed loop is sufficiently stable lies not in technical barriers, but in the fact that it is a business model that has been repeatedly verified and is self-consistent.

Search ads are settled by clicks, information feed ads measure value by metrics such as dwell time, e-commerce platforms take commissions based on transaction volume, and local life services share revenue based on in-store redemption—the premise for all monetization links to hold is that users complete the full interaction chain within the app.

The platform's control over users' interaction paths is essentially the foundation of its commercial pricing power. Whoever controls users' browsing paths, decision-making processes, and transaction entry points possesses the dominant power in distributing the entire value chain.

Technology can be caught up with, and product features can be imitated, but the traffic distribution power accumulated from more than a decade of user habits constitutes the commercial foundation of the mobile internet. Any technology that attempts to shake the premise that users complete full-link actions within apps may impact the foundation of the entire value chain.

The emergence of the SAEP protocol is a consultation mechanism built by phone manufacturers in response to this commercial reality.

It no longer has phone manufacturers unilaterally decide screen reading and automation permissions, but instead establishes an ex ante declaration mechanism: returning to developers the decision-making power over whether to allow AI assistants to perform screen automation operations on their own applications. Developers can independently define permission boundaries and set operation restrictions, and even after the public comment period ends, they can submit refusal declarations at any time.

The design idea of this mechanism is similar to robots.txt on the PC internet: developers independently define accessible boundaries and then explore the possibility of cooperation. But the premise for the SAEP mechanism to take effect is that developers are willing to cooperate.

Judging from the results of the first public notice, the degree of cooperation from leading platforms is far lower than industry expectations. Behind this collective refusal, the target is far from merely a single phone manufacturer: super apps in different tracks all base themselves on their own commercial foundations, weigh the costs and risks of opening permissions, have their own commercial considerations, and each refusal has corresponding practical reasons.

The shift of decision entry points leaves platforms facing a role change from front end to back end

The core concern of apps is not distrust of AI technology itself, but the lack of a compensation mechanism for the commercial losses brought by opening permissions.

The core demand of social and gaming platforms is to hold on to user time and immersive interaction, while AI acting on users' behalf reduces their willingness to actively open, browse, and interact with apps; the user time on which feed advertising depends risks being diverted. For social and gaming ecosystems that rely on user time and advertising monetization, this could hit their commercial foundation.

The focus of e-commerce and local life platforms is on protecting the transaction chain and the power to allocate traffic. If AI agents intervene deeply, cross-platform price comparison and automatic ordering are expected to become a potential trend, and the search ranking, ad slots, and store traffic allocation mechanisms that platforms have carefully designed may be bypassed in the AI execution chain.

If they lose control over users' decision paths, the foundation of platform advertising and commission revenue risks being shaken.

The concerns of large-model vendors are essentially mutual fortification among similar products. AI products are naturally in direct competition; if competitor agents are allowed to operate automatically within one's own apps, user behavior and conversation data may be indirectly collected by rivals, and product capability boundaries risk being probed.

In the large-model race, where data is a core asset, such openness is akin to exposing one's cards to competitors.

Financial apps' concerns stem from hard constraints at the risk-control and compliance levels. In screen automation scenarios, transaction liability allocation, identification of abnormal transactions, and the corresponding regulatory rules still lack clear standards. For financial institutions, fund safety and compliance take far higher priority than experience and efficiency gains; before the relevant rules are clear, closing permissions is the lowest-risk choice.

It is not hard to see that the root of refusing openness is not that AI technology itself is unreliable, but that platforms anticipate their core interests may be harmed after opening up, and there is no corresponding compensation mechanism. Social platforms worry about losing user time, e-commerce worries about the transaction chain being bypassed, large-model vendors fear data leaks, and financial institutions prioritize guarding against security risks.

The real impact of GUI-automation AI Agents on the existing ecosystem is not that AI clicks the screen on users' behalf, but that it may fundamentally drive a migration of users' decision entry points.

The ideal scenario depicted by GUI automation is this: a user gives a natural-language instruction, and AI completes an entire task flow across apps—price comparison, ordering, booking, queries, and so on. Users need not open each app one by one, nor browse homepage recommendations; native ads may lose their chance to reach them, and the search traffic allocation mechanisms platforms painstakingly built may be weakened in the AI execution chain.

For platforms, this means a role change: from a business front end that directly serves users to a back-end service invoked by AI. The front end holds traffic, ad slots, and pricing power; once turned into the back end, it can only passively respond to invocation requests, and how to realize revenue is a difficult problem many platforms must confront.

More alarming is that once such user habits form, the cost of reversing them later will rise significantly.

The logic of platforms' choices thus becomes clear: rather than waiting for AI agents to gradually divert users' decision-making steps, it is better to draw boundaries in advance. Refusing GUI automation is not resistance to technological progress, but a defense of their most core asset—the dominant power over user interaction.

Computing power has room for continuous optimization and upgrading, but the user habits, traffic allocation power, and transaction chains that super apps have accumulated over more than a decade will not automatically restructure with technological iteration.

A reverse game may even emerge: the stronger the Agent's capabilities and the deeper its penetration into scenarios, the stronger platforms' wariness, and ecosystem barriers may keep hardening.

The second half of AI Agents: the tech race ends, the game of interests begins

The all-purpose Agent narrative once highly anticipated by the industry is being continuously revised in the face of ecosystem reality.

The first route is GUI screen automation, i.e., the approach corresponding to the SAEP protocol. Its advantages are intuitive: no need for apps to develop separate interfaces, a low barrier to implementation, and in theory coverage of a large number of application scenarios.

But its shortcomings are equally prominent: this capability depends heavily on the willingness of apps to cooperate, and a single statement from a platform can cut off the function; a page redesign may cause recognition to fail, and a series of issues such as security audits and allocation of rights and responsibilities remain unsettled to this day.

This route is better suited to long-tail apps willing to open permissions, and is unlikely to become a universal solution for connecting leading super apps.

The second route is the native API collaboration model: AI assistants invoke platform services through standardized interfaces. This approach offers stable interaction, controllable permissions, and clear rights and responsibilities; platforms can choose the scope of capabilities to open and design revenue-sharing rules on their own.

The industry is more likely to move toward an intermediate form: some apps with an open attitude open native interfaces, and AI takes on high-value tasks in safe and stable scenarios; for leading super apps that refuse openness, AI can only identify task blockers, prompt users to intervene manually, and after the jump is completed, no longer automatically execute subsequent actions.

The all-purpose Agent narrative of settling all matters in one sentence will likely see its applicable boundaries continuously compressed by ecosystem reality. Phone manufacturers have often depicted their product vision as cross-app, full-chain operation; after the SAEP incident, this narrative needs recalibration: the obstacle comes not from technical capability, but from constraints at the ecosystem level.

For a phone AI Agent to intervene deeply in users' full-chain operations, three core parties are indispensable: users pursue efficiency, AI vendors hope to expand capability boundaries and tap commercial value, and platforms must defend their dominance over traffic and transactions. The demands of the three parties are inherently in tension, and the existing protocol framework has only established basic rules without solving the most core problem of interest distribution.

Platforms opening permissions is essentially giving up part of their dominance over user decisions; without a commensurate revenue-sharing, traffic-exchange, or value-sharing mechanism, such a transfer of permissions lacks commercial rationality.

Over a longer time horizon, the focus of competition in the second half of phone AI Agents may no longer be technical metrics such as model parameters and computing power, but whether a collaboration scheme acceptable to all parties can be found within the existing traffic closed-loop system.

Technology can build an agent that understands the screen and executes instructions, but making it stably handle affairs for users in the real mobile ecosystem is essentially a commercial proposition: what do platforms gain by opening this door of permissions?

The computing power race will continue, but what determines the ultimate ceiling of phone AI Agents is no longer computing power itself.

The mobile internet has never built technical walls, but commercial walls. As long as the problem of revenue distribution behind the walls lacks a solution, even the strongest computing power will find it hard to penetrate the traffic closed loops of super apps.