文｜硅碳变量，作者｜胡一刀
10月14日，豆包手机助手官方公众号发文称，《屏幕自动化操作声明协议》（SAEP，Screen Automation Execution Protocol）30天公示期已结束，平台将逐步放开第三方应用的屏幕自动化操作权限。
据公示，最新拒绝操作的应用清单涵盖拼多多、Kimi、快手系列应用、美团系列应用、小红书、DeepSeek、中国工商银行等，以及淘宝、千问、支付宝等23款阿里巴巴及关联生态应用，微信、QQ、王者荣耀等65款腾讯及关联生态应用，共计151款。
据此前介绍，SAEP是面向第三方App的AI操作规则，开发者可自主声明是否允许AI助手对自身应用执行屏幕自动化操作，划定操作边界与限制范围。官方强调，公示期结束后，应用开发者仍可随时提交拒绝声明，收到后豆包手机助手将及时关闭相关操作。
超级App的根基，藏在流量闭环之中
手机AI Agent的算力焦虑曾是行业两年多来的核心叙事，而它消散的速度超出了不少人的预期。
短短两年时间，端侧大模型参数规模实现了从数十亿到上百亿的跨越，旗舰芯片NPU算力密度实现大幅提升；内存带宽瓶颈、量化压缩算法、端侧推理框架等底层工程难题，在芯片厂商与手机厂商的联合攻坚下持续取得进展。
如今不少旗舰机型已经能够在本地承载语音交互、意图理解、图像识别这类基础AI能力，多模态模型对手机屏幕的感知也不再局限于识别按钮坐标，开始具备解析页面浅层语义的能力。
技术迭代的脚步仍在加速，但一个此前长期被行业低估的核心矛盾正迅速浮出水面：当AI初步具备看懂屏幕、模拟点击、跨应用拆解执行指令的能力时，决定手机AI Agent边界的早已不再是技术上限，而是应用生态的准入红线。
算力定义了AI Agent的能力天花板，而生态规则划定了它的实际活动边界。
二者的难度完全不在一个维度：算力瓶颈依托资本投入与技术迭代，存在持续改善的空间；但生态秩序的重构，撬动的是移动互联网沉淀十余年的商业利益底盘。
要读懂这批App集体关闭权限背后的动因，需要回到移动互联网运行十余年的核心商业模型：流量闭环。
过去十余年，移动互联网平台的价值链条始终围绕这套闭环体系运转：平台以内容、社交关系、商品供给、本地服务为锚点，将用户时长与注意力沉淀在自有App生态内，再通过信息流广告、搜索竞价、交易佣金、会员订阅等方式完成价值变现。
用户在App内的每一次浏览、点击、比价、下单都是流量价值的载体；广告主为曝光付费，商家为流量付费，用户为服务付费，平台则在每一层价值交换中抽取收益。
这套流量闭环之所以足够稳固，核心不在于技术壁垒，而在于它是一套被反复验证和自洽的商业模型。
搜索广告按点击结算，信息流广告以停留时长等指标衡量价值，电商平台按成交额抽佣，本地生活按到店核销分成——所有变现环节成立的前提，都是用户在App内完成完整交互链路。
平台对用户交互路径的控制权，本质上是其商业定价权的根基。谁掌握了用户的浏览路径、决策过程与交易入口，就拥有整条价值链的分配主导权。
技术可以被追赶，产品功能可以被模仿，但十余年用户习惯沉淀下来的流量分配权，构成了移动互联网的商业底盘。任何试图动摇用户在App内完成全链路动作这一前提的技术，都有可能冲击整条价值链的根基。
SAEP 协议的出现，是手机厂商针对这一商业现实搭建的协商机制。
它不再由手机厂商单方面决定屏幕读取与自动化权限，而是建立了一套事前声明机制：把是否允许AI助手对自身应用执行屏幕自动化操作的决定权交还给开发者。开发者可自主划定权限边界、设置操作限制，即便公示期结束，也能随时提交拒绝声明。
这套机制的设计思路类似PC互联网的robots.txt，由开发者自主划定可访问边界，再探讨合作可能性。但SAEP机制能够落地生效的前提，是开发者愿意配合。
从首批公示的结果来看，头部平台的配合度远低于行业预期。这场集体拒绝的背后，远非单纯针对某一手机厂商：不同赛道的超级App，都立足自身商业底盘，权衡开放权限的成本与风险，各有商业考量，拒绝都有对应的现实理由。
决策入口位移，平台面临前台到后端的角色变局
App顾虑的核心并非不信任AI技术本身，而是开放权限带来的商业损失缺少补偿机制。
社交与游戏平台的核心诉求是守住用户时长与沉浸式交互，而AI代操作降低了用户主动打开、浏览、交互应用的意愿，信息流广告赖以生存的用户时长，存在被分流的风险。对于依靠用户时长与广告变现的社交、游戏体系而言，这可能冲击其商业基本盘。
电商与本地生活平台的重点，在于守护交易链路与流量分配权。若AI代理深度介入，跨平台比价和自动下单有望成为一种潜在趋势，平台精心设计的搜索排序、广告展位、店铺流量分配机制，有可能在AI执行链路中被绕开。
倘若失去对用户决策路径的掌控，平台广告与佣金收入的根基便存在动摇的风险。
大模型厂商的顾虑，本质是同类产品之间的相互设防。AI产品天然存在直接竞争，若允许竞品智能体在自家应用内自动操作，用户行为、对话数据存在被对手间接采集的可能，产品能力边界也有被探测的风险。
在数据是核心资产的大模型赛道，这类开放行为近似于向竞争对手暴露自身底牌。
金融类App的顾虑，源于风控与合规层面的硬性约束。屏幕自动化场景下，交易责任划分、异常交易识别、对应的监管细则，目前仍缺少清晰标准。对金融机构而言，资金安全与合规优先级远高于体验效率提升，在相关规则明确前，关闭权限是风险最低的选择。
不难看出，拒绝开放的根源不在于AI技术本身不可靠，而是平台预判开放后核心利益可能受损，且缺少对应的补偿机制。社交平台担忧时长流失，电商担心交易链路被绕过，大模型厂商忌惮数据泄露，金融机构首要防范安全风险。
GUI自动化AI Agent对现有生态的真正冲击，不在于AI代为点击屏幕，而在于它可能从底层促成用户决策入口的迁移。
GUI自动化所描绘的理想场景是：用户给出一句自然语言指令，AI跨应用完成比价、下单、预约、查询等一整套任务流程。用户无需逐个打开App，也不会浏览首页推荐内容，原生广告或将失去触达机会，平台苦心搭建的搜索流量分配机制，在AI执行链路中可能被弱化。
对平台而言，这意味着角色变迁：从直接承接用户的业务前台，转变为被AI调用的后台服务。前台掌握流量、广告位与定价权，转为后台之后，只能被动响应调用请求，收益如何兑现是不少平台需要直面的难题。
更值得警惕的是，一旦形成这类用户习惯，后续扭转的成本会显著抬升。
平台的选择逻辑由此变得清晰：与其等待AI智能体逐步分流用户决策环节，不如提前划定边界。拒绝GUI自动化并非抵触技术进步，而是守住自身最核心资产——用户交互的主导权。
算力拥有持续优化升级的空间，但超级App历经十余年沉淀的用户习惯、流量分配权与交易链路，不会随技术迭代自动重构。
甚至会形成反向博弈：Agent能力越强、渗透场景越深，平台的戒备心理越强，生态壁垒或将持续加固。
AI Agent下半场：技术竞赛落幕，利益博弈登场
曾经被行业寄予厚望的全能Agent叙事，正在生态现实面前被不断修正。
第一条路线便是GUI屏幕自动化，也就是SAEP协议对应的方案。它的优势很直观，无需 App单独开发接口，落地门槛低，理论上可覆盖大量应用场景。
但其短板同样突出，这套能力高度依赖应用方的配合意愿，平台一纸声明即可切断功能；页面改版就可能造成识别失效，安全审计、权责划分等一系列问题至今没有定论。
这条路线更适配愿意开放权限的长尾应用，很难成为打通头部超级App的通用解法。
第二条路线为原生API协作模式：AI助手依托标准化接口调用平台服务。该方案交互稳定、权限可控、权责清晰，平台可自主选择开放能力范围，自主设计分润规则。
代价则是推进节奏缓慢，需要逐一对接商务、敲定接口规范、协商收益分成。而多数头部平台本身已有AI业务布局，未必愿意向外部AI助手开放核心接口。
行业更有可能走向中间形态：部分持开放态度的应用开放原生接口，AI在安全稳定场景中承接高价值任务；面对拒绝开放的头部超级App，AI仅能识别任务卡点，提示用户手动介入，完成跳转后不再自动执行后续动作。
一句话搞定全部事务的全能Agent叙事，其适用边界或将被生态现实持续压缩。手机厂商此前常以跨应用全链路操作描绘产品愿景，SAEP事件之后，这套叙事需要重新校准：阻碍并非来自技术能力，而是生态层面的约束。
手机AI Agent若要深度介入用户全链路操作，离不开三方核心主体：用户追求使用效率，AI厂商希望拓展能力边界、挖掘商业价值，平台则要守住流量与交易的主导权。三方诉求天然存在张力，现有的协议框架仅搭建起基础规则却没能解决最核心的利益分配难题。
平台开放权限本质是出让部分用户决策主导权，倘若没有对等的分润、流量置换或是价值共享机制，这种权限让渡便缺少商业层面的合理性。
拉长时间维度来看，手机AI Agent下半场的竞争重心或将不再是模型参数、算力这类技术指标，而是能否在既有的流量闭环体系内找到多方可接受的协作方案。
技术可以打造读懂屏幕和执行指令的智能体，但想要让它在真实移动生态里稳定为用户处理事务，本质是一道商业命题：平台打开这扇权限之门，能够收获什么？
算力竞赛仍会持续，但决定手机AI Agent最终发展上限的早已不是算力本身。
移动互联网筑起的从来不是技术围墙，而是商业围墙。只要围墙背后的收益分配问题缺少解决方案，再强大的算力也难以穿透超级App的流量闭环。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It's a vinyl silicon carbon variable.
On 14 October, the Doubao mobile phone assistant, official WeChat Official Account, wrote that the 30-day public announcement period for the " Statement of Action on Screen Automation " (SAEP, Screen Automation Exchange Production) had ended and that the platform would gradually liberalize the third-party application of screen automation permission.
According to the public announcement, the most recent list of applications for refusing to operate covers a wide range of applications, Kimi, Express Series applications, American Series applications, REDnote, DeepSeek, China Chamber of Commerce and Industry, as well as 23 Alibaba and associated eco-applications such as WeChat, QQQ, Wang Glory and others, totalling 151.
As described above, SAEP is an AI operating rule for a third party, App, and the developer is free to declare whether or not AI's assistant is allowed to automate his/her own application on the screen, delimiting the operational boundaries and limits. The official stressed that the application developers could still submit a rejection statement at any time after the announcement period had ended and that, upon receipt, Doubao mobile phone assistants would close the operation in time.
Super App's foundation, hidden in a traffic ring.
The calculus anxiety of the mobile phone AI Agent was the core narrative of the industry for more than two years, and it spread faster than many expected.
In a short period of two years, the end-side LLM parameter size has been crossed from billions to billions, and the base NPU calculator density of the flagship chip has been significantly increased; lower-level engineering challenges, such as memory bandwidth bottlenecks, quantization compression algorithms, end-side reasoning frameworks, have continued to advance under the combined attack of chip and mobile phone manufacturers.
Many flagship aircraft are now capable of carrying local basic AI capabilities such as voice interaction, understanding of intent, image recognition, and the perception of mobile screens in multi-modular models is no longer limited to identifying button coordinates and is beginning to have the capability to decipher the subsonics of the page.
The pace of technology overlaps is accelerating, but a core contradiction that was previously underestimated by the industry is rapidly emerging: when AI has the initial ability to read screens, simulate clicks, and cross-apply dismantling to implement directives, the determination of the AI Agent border is no longer a technical cap, but rather an application of ecological red-line access.
The algorithm defines the AI Agent capacity ceiling, while ecological rules delineate the actual boundaries of its activities.
The difficulty of the two is far from one dimension: arithmetical bottlenecks depend on capital inputs and technology overlaps, and there is room for continuous improvement; however, the ecological order is re-engineered by a chassis of commercial interests that has been deposited in the mobile Internet for more than a decade.
To understand the driving forces behind the collective closure of permission by App, it is necessary to return to the core business model that has been running on the mobile Internet for more than a decade: traffic closure.
Over the past decade or so, the value chain of the mobile Internet platform has been organized around this closed ring system: the platform is anchored in content, social relationships, supply of goods, local services, placing users' time and attention within their own App ecology and completing value realization through information stream advertisements, search bids, trade commissions, membership subscriptions, etc.
Each browsing, clicking, bargaining, and downlist of users in the App is a carrier of flow value; the advertiser is exposed to paid, the commercial is exposed to paid, the user is serving paid, and the platform extracts the proceeds from each layer of value exchange.
It is not a technical barrier, but rather a repeatedly validated and self-conforming business model.
Search advertisements are measured by clicking, information stream advertisements by indicators such as length of stay, electricians' platforms are paid off, local lives are split by shop-to-shop write-off - all realizations are premised on complete interactive links being completed by users in App.
The platform ' s control over the user ' s interactive path is essentially the foundation of its commercial pricing rights. Those who have access to users ' browsing paths, decision-making processes and trading outlets have the distributional ownership of the whole value chain.
Technology can be pursued and product functions can be imitated, but the flow distribution power that users have been accustomed to settling for more than a decade constitutes a commercial chassis of mobile Internet. Any technology that attempts to undermine the premise that users complete full-link operations within the App has the potential to strike the roots of the entire value chain.
The emergence of the SAEP agreement is a consultative mechanism established by mobile phone manufacturers to address this commercial reality.
Instead of unilaterally deciding on screen reading and automation of permission by mobile phone manufacturers, it established a system of ex ante declarations: handing over to developers the decision whether to allow AI ' s assistant to automate his own application. The developer is free to demarcate the permission boundary and set operational limits, and can submit a declaration of rejection at any time even after the publicity period has ended.
The mechanism is designed along the lines of the robots.txt of the PC Internet, where the developers are autonomous in delineating the accessible borders and exploring possibilities for cooperation. But the prerequisite for the SEPA mechanism to come into effect is the willingness of developers to cooperate.
From the results of the first public announcements, the alignment of the head platform is far below industry expectations. This collective rejection is far from being directed solely at a mobile phone manufacturer: the SuperApp of different tracks is based on its own commercial chassis, weighing the costs and risks of opening up permission, each of which has commercial considerations, and there are realistic reasons for rejection.
Transfer of the decision entry and the platform faced a change of character from the front to the back
The core of App ' s concern is not the lack of trust in AI technology per se, but the lack of compensation mechanisms for the commercial losses caused by the opening of permission.
The core claim of the social and play platform is to keep user time and immersion interactive, while AI-based operations reduce the user ' s willingness to open, browse and interact, and information stream advertising is based on long-term users who risk diversion. For social, game-based systems that rely on user time and advertising for their realization, this may impact their business fundamentals.
The focus of the electrician and local life platform is on safeguarding the trading chain and distribution of traffic. With the deep involvement of AI agents, cross-platform pricing and automatic listings are expected to be a potential trend, and the platform ' s carefully designed search sequencing, advertising booths, and shop traffic distribution mechanisms may be bypassed in the AI implementation link.
If there is a loss of control over the user ' s decision path, there is a risk that platform advertising and commission revenues will be shaken.
The concerns of LLM manufacturers are essentially mutual defence between the same products. There is natural direct competition for AI products, and there is a risk that the boundary of product capacity will be detected if the competition AI agent is allowed to operate automatically in its own home application, and that user behaviour, dialogue data may be collected indirectly by the counterparty.
In the LLM track, where the data are the core asset, such openness is close to exposing itself to competitors.
The concern of the financial type App stems from the rigidity of controls and compliance. Clear criteria for the allocation of liability for transactions, the identification of unusual transactions and the corresponding regulatory rules are still lacking in an automated screen scenario. For financial institutions, financial security and compliance priorities are much higher than experienced efficiency gains, and closing permission is the least risky option until the relevant rules are clear.
It is easy to see that the reason for the refusal to open up is not that AI technology itself is unreliable, but rather that the platform predicts that core interests may be compromised after opening up and that there is a lack of corresponding compensation mechanisms. Social platforms are worried about the loss of time, electricians are worried about bypassing the trading chain, LLM producers are afraid of the disclosure of data, and financial institutions are the first to protect against security risks.
The true impact of GUI automated AI Agent on existing ecology lies not in the fact that AI is clicking on screens, but in the fact that it can contribute to the migration of user decision-making portals from the bottom.
The ideal scenario for the GUI automation is that users give a natural language instruction and AI cross-apply a set of task processes such as value, order, appointment, query, etc. Users do not have to open the App and do not browse the recommendations on the front page, either the original ad or the lost touch, and the search flow allocation mechanism built up by the platform may be weakened in the AI implementation link.
For the platform, this means changing roles: from a business front to a back-office service directly to the user. Once the front desk handles traffic, advertising positions and pricing rights and moves backstage, it can only respond passively to calls, and the realization of the benefits is a challenge that many platforms need to face directly.
It is even more alarming that, once such user habits are developed, the cost of subsequent reversals will increase significantly.
The logic of the platform ' s choice became clear: rather than waiting for AZAI agent to gradually divert the user ' s decision-making chain, it would be better to delineate the boundary in advance. The rejection of GUI automation is not a violation of technological progress, but rather a defence to its core asset, the interactive ownership of users.
The algorithm has room for continuous optimization, but SuperApp does not automatically re-engineer with technology over a decade of settled user habits, flow allocation rights and trading links.
It can even create a reverse game: the stronger the Agent's capacity, the deeper the permeability, the greater the platform's alertness, the ecological barriers may continue to be reinforced.
The second half of AI Agent: the end of the technology competition, the end of the profit game
The All-powerful Agent narrative, which had been placed in high industry expectations, was constantly being revised in the face of ecological realities.
The first route is to automate the GUI screen, which is the program for the SAEP protocol. Its advantages are intuitive, it does not require the separate development of the interface by App, it has a low landing threshold and theoretically covers a large number of applications.
However, its shortboard is equally prominent, and the capability is highly dependent on the willingness of the operator to cooperate, and the platform can be cut off by a paper statement; a re-version of the page may result in a failure of identification, and a range of issues, such as security audits and the allocation of responsibilities, remain open.
This route is more amenable to the long-tail applications of permission, and it is difficult to become a universal solution to head-to-head superapp.
The second route is the original API collaboration model: AI assistant calls Platform services based on standardized interfaces. The programme is interactively stable, permission is manageable and has clear authority and responsibility, and the platform has the autonomy to choose the scope of open capacity and to design the rules of decentralization.
The cost is slow to advance the pace, requiring one-on-one business, finalization of interface specifications and negotiation of the share of proceeds. Most of the heads of flat Z89Z have an AI operating layout and are not necessarily willing to open core interfaces to external AI assistants.
Industry is more likely to move towards an intermediate pattern: partially open-minded application of open primary interfaces, with AI taking on high-value tasks in a secure and stable scenario; in the face of a refusal to open the head super App, AI can only identify the taskcard points, prompting users to intervene manually and no automatic follow-up actions after jump.
The full-fledged Agent narrative of a single sentence of all things is either applied at a boundary or will be continuously compressed by ecological realities. Mobile phone manufacturers had previously used the application of full-chain operations to depict product visions, and after the SAEP event, the narrative needed to be re-corrected: the impediment was not from technical capacity, but from ecological constraints.
Mobile phone AI Agent is required to be deeply involved in the user ' s full-link operation: the user seeks efficiency, the AI manufacturer wants to expand the capacity boundaries and exploit commercial values, and the platform has to maintain ownership of traffic and transactions. There is a natural tension in the tripartite claims, and the existing framework of agreements is based only on foundational rules that fail to resolve the central distribution of benefits.
The platform ' s opening to permission is essentially a decision-making power for some users, and the permission concession lacks commercial legitimacy if there is no equivalent differentiation, traffic replacement or value-sharing mechanism.
For long periods of time, the focus of competition in the second half of the mobile phone AI Agent may no longer be technical indicators such as model parameters, calculus, but rather the ability to find a multi-acceptable collaborative solution within an established traffic closed ring system.
Technology can create a AI agent that reads the screen and executes the command, but it is essentially a commercial proposition to stabilize it in a real mobile environment for users: The platform opens the door to permission. What can it yield?
The calculus competition will continue, but it is no longer the calculus itself that determines the ultimate upper limit for the development of the mobile phone AI Agent.
Mobile Internet has never been built as a technical fence, but as a commercial fence. As long as there is no solution to the distribution of the proceeds behind the wall, it will be difficult to penetrate the traffic closure of SuperApp.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.