Skip to content
DCC · DATA COMPLIANCE CHINA China data law, for overseas counsel.
§ 121 · AI-AGENTS

The Agent Authorization Standoff After SAEP: Li Wenlong on Why 155 Apps Refused the Doubao Phone Assistant, and the Case for Reasonable, Non-Discriminatory, Reciprocal Interoperability Rules

A translation of Li Wenlong's October 26, 2026 essay on the Screen Automation Execution Protocol (屏幕自动化操作声明协议, SAEP): how the protocol works, why technical trust was never the whole problem, three kinds of refusal, and why mandatory opening is the wrong tool

On October 14, 2026 the Doubao phone assistant announced that the 30-day public notice period for its Screen Automation Execution Protocol (SAEP) had ended, and published a list of 155 apps that had refused GUI automation: 66 from Tencent's ecosystem, 23 from Alibaba's, plus Meituan, Kuaishou, Xiaohongshu, banks and AI apps. Li Wenlong (李汶龙, 科技利维坦) asks whether they refused because the protocol is unsound or because of the politics of agent authorization. He walks through SAEP's design: policy declared in AndroidManifest.xml and through a ContentProvider, three scopes (App, Activity, Intent), seven restriction fields and four system decisions (ALLOW, BLOCK, CALL_USER, CLARIFY). He finds it a real access-control scheme whose weak point is that rule-making, interpretation and enforcement sit with one party. He then reads the refusals through the Prisoner's Dilemma, the Stag Hunt and asymmetric games, sorts refusals into three kinds (security and compliance, structural distrust, and security used as a pretext against competitors), argues that anti-monopoly-style mandatory access is unjustified at this stage, and proposes a floor of shared rules built on reasonableness, non-discrimination and functional reciprocity. The essay predates MIIT's Notice 工信厅信管〔2026〕42号, which has since made app developers' declared consent a condition of agent invocation.

Editor’s Note — DCC.

Why this is on DCC. The essay is about a private protocol, not a statute. DCC runs it because the question it works through, who decides whether an AI agent may operate someone else’s app, is the one MIIT answered eleven days later. Part I, item (4) of the MIIT Agent Terminal Notice (工信厅信管〔2026〕42号, published November 6, 2026) provides that “without a declared consent, an agent shall not forcibly invoke or call up the app”. SAEP, the protocol analyzed here, worked the other way: silence after a 30-day notice period counted as permission. This essay is the best available account in Chinese of what that protocol was, how the market answered it, and what its author thinks should come instead. The essay was published on October 26, 2026 and does not discuss the Notice.

Sequence. It follows Li’s two earlier pieces, both on DCC: Reverse Interoperability (December 2025) and AI Agents and the Limits of Consent (January 2026). The “consensus declaration” announced at the end was released on November 7, 2026; DCC’s translation is in The Initiative on Basic Rules for Agent Authorization and Interoperability.

What is translated. The essay is translated in full, with one exception: three passages recounting the textbook history of the game theory models are condensed, and marked where they are. The four diagrams from Doubao’s developer documentation are reproduced with DCC’s English reading. Section headings are the author’s. Market facts are as the author reports them from Chinese press coverage.

— Mandatory opening is unreasonable, and letting apps shut out competitors in the name of security is not the best answer either. We need to build consensus step by step, from the bottom up.

By Li Wenlong (李汶龙)

On October 14, 2026, the Doubao phone assistant announced that the 30-day public notice period for its Screen Automation Execution Protocol (屏幕自动化操作声明协议, SAEP) had ended, and that it intended to open up screen-automation permissions for third-party apps step by step.

Published at the same time was a list of refusals. As of mid-October, 155 apps had expressly refused automated operation by the Doubao phone assistant through the graphical user interface (GUI): 66 apps from Tencent and its affiliated ecosystem, 23 from Alibaba and its affiliated ecosystem, and also Meituan, Kuaishou, Xiaohongshu, banks and a number of AI apps. These 155 apps cover, more or less, the great majority of the internet companies, banks and telecom operators we deal with every day. They brought the question of authorization in the age of agents into public view for the first time in so concentrated a form.

What is intriguing is who did not expressly refuse. Li Ling (2026), a reporter with Nandu AI Outpost (南都AI前哨站), found in hands-on testing that the Doubao phone assistant could complete part of a purchase on JD.com, and apps such as Ctrip, Tongcheng, Weibo, iQIYI and Bilibili were also observed to be operable. Looking further, most of the apps that did not refuse have strategic cooperation with ByteDance products: JD Logistics with Douyin instant retail, iQIYI with Douyin, and so on. More interesting still is Baidu. According to Caijing (2026), most Baidu apps do not appear on the refusal list, but Baidu Maps expressly refused. A Baidu insider interviewed by Caijing answered that shipments of the Doubao phone are small and it is not much of a problem for Baidu.

What does ByteDance want to do? And why are the other players in the ecosystem unwilling to take part?

There are several dimensions behind a refusal to interoperate. The one mentioned most is security defense (because the concept covers everything): automated operation may break existing risk controls, transactions and services. The technical preview of the Doubao phone assistant, launched in December 2025, is a typical example. The very next day it was restricted by many apps on grounds of account security, human verification and risk. WeChat, Alipay, Meituan, Pinduoduo (PDD) and others put the Doubao phone assistant on their restricted lists one after another.

But the reasons for refusing to interoperate are, as set out below, more complicated. In the end this is a fight for survival in the mobile internet’s period of transition. The agent is not simply a new technology that has dropped into the existing app ecosystem from the sky. Set in the current political economy of the internet, the agent should be understood as competitive infrastructure that replaces the existing app ecosystem and becomes the next generation’s entry point for traffic and transactions.

Some key facts that the public narrative tends to blur. First, 155 apps does not mean 155 companies. Second, “refusing access” means refusing screen automation by a particular agent; users can still install, open and manually use these apps on the phone. Third, that an app, or the ecosystem behind it, has not taken a position does not mean it approves and permits. And an app that has taken a position will not necessarily go on refusing.

In essence the Doubao phone assistant has taken the opt-out path. Its legitimacy comes from openness and transparency and from giving the other players in the ecosystem time to respond. In theory, some app may later come forward to say it never saw the declaration and was invoked without authorization. Judging by the market’s reaction, though, SAEP received a great deal of attention and immediate feedback. Talking with members of my group chat, I was struck by an interesting pattern. Over the decades in which the internet rose and its regulatory rules were built, the positioning and default state of the law have determined the path and extent of technological development. From the early consent regime (the opt-out debates around the e-Privacy Directive at the turn of the century), to the Robots.txt protocol, to the text-and-data-mining exception introduced by the EU copyright reform, to the legitimacy of training data today: in every case the balance between technological development and the protection of rights was struck by way of opt-out. But opt-out is only the use, and abuse, of a default. This is an environment of rapid market fusion, and also one of long-run, dynamic gaming.

To understand how the SAEP affair came about, one has to go back to December 2025 (see The Doubao Phone Assistant Is, at Bottom, a Problem of Passive Interconnection), when the technical preview of the Doubao phone assistant met restrictions from some apps (for various reasons, including account security, identity authentication and automated risk control). GUI technology combined with cooperation from a handset maker had made automated operation by an agent work, and opened a new traffic entry point. In the foreseeable future, therefore, apps may gradually retreat to the background while users organize cross-platform tasks through agents.

After that, Doubao tried to turn access relationships that had depended on ad hoc coordination between products into rules that can be publicly declared and enforced by the system. Against this background the Doubao phone assistant launched SAEP and its notice procedure in mid-September. During the notice period it adopted a fairly cautious arrangement of not operating by default. After the notice period, operating capability is opened step by step for apps that have not declared a refusal, though an app may still express refusal at any time. So in essence this is an opt-out logic: participation by default unless refusal is actively expressed (as distinct from opt-in: non-participation by default unless willingness is actively expressed). [DCC: the Chinese text reads “opt in” at the start of this sentence; the definition that follows and the rest of the essay show that opt-out is meant.]

Some describe SAEP as the Robots protocol of the AI era. Logically the comparison is not rigorous. Still, it deserves a few more words, because it helps to show the logic involved. Both concern the boundary of access; in both, one system tries to connect with, or “shake hands” with, another; and both allow the provider of the resource or service to express refusal. The traditional Robots Exclusion Protocol is mainly directed at web crawlers fetching website resources, and compliance depends to a considerable degree on the crawler’s operator. SAEP faces GUI agents that can act, and the conduct involved may directly change the state of an account, send messages, delete content or even trigger a transaction. The difference in what is authorized, and in its effects, is plain.

Another key difference is who initiates the protocol. The Robots protocol is made by the website (the party being accessed), allowing or prohibiting crawlers and setting their scope; if the website has made no protocol, it may be crawled by default. SAEP is initiated by the Doubao phone assistant (the party seeking access), which invites the apps and ecosystems concerned to refuse actively; if an app does not refuse, it is understood by default to be invocable. If the Robots protocol had websites take defensive measures for the sake of the internet economy, SAEP is an “offensive” measure taken by ByteDance alone to bring its agent business into being. Both, however, rest on an opt-out architecture, built on certain expectations or duties of conduct on the part of the accessed party and on procedural safeguards, where joint action is hard to achieve.

This essay attempts a preliminary analysis of two questions.

First, a basic judgment. Did the great majority of the ecosystem refuse because ByteDance’s own scheme is unreasonable or one-sidedly forceful, or is the technical scheme itself broadly sound, with the refusals explained by the political game behind agent authorization? The former calls for a close look at SAEP’s architecture; the latter calls for some game-theoretic thinking.

Second, from the perspectives of market negotiation, law and regulation, how should this mass refusal be understood, and what does agent authorization actually need now (coordination, trust, or regulatory intervention) for the industry to move forward in a healthy way?

The micro-structure of SAEP

SAEP’s basic operating mechanism is to convert a third-party app’s declaration of restrictions on agent operation into an access policy that the terminal system can recognize and enforce. As the diagram in the integration documentation shows, an app first declares in AndroidManifest.xml whether it adopts SAEP, and supplies the specific access policy through a ContentProvider. The policy management module on the device discovers, fetches and manages these declarations. When an agent attempts to operate an app, the system makes an access decision by reference to the relevant policy and enforces the restriction through the terminal system’s control capability. An app can thus express which operations may proceed and which should be prohibited or handed over to the user, and the agent’s actual operation is bound by the system’s policy ruling. The key to this architecture is that the app declares the access boundary and the terminal system interprets and enforces the rule. SAEP thereby converts access restrictions that were scattered across different apps into one fairly uniform technical mechanism. But how reliable policy interpretation, enforcement and verification are also becomes an important factor in whether third-party apps will trust it.

Figure 1 | How the SAEP protocol works
Flow diagram in Chinese with four blocks. App integration: a static declaration of the SAEP policy in AndroidManifest.xml, a ContentProvider for dynamic policy updates and queries, and a ContentProvider for querying agent operation logs. Agent policy and management: a system service that queries the latest policy periodically or on demand. Agent log management: a log service that stores and manages logs. Pre-execution check: the agent queries the policy before starting an operation, and a system service verifies the agent's identity and returns the final policy.

Diagram from the original post (source: Doubao's SAEP integration documentation), in Chinese. DCC's English reading follows.

DCC’s English reading. App integration: a static declaration of the SAEP policy in AndroidManifest.xml; a ContentProvider for dynamic policy updates and a query interface; a ContentProvider through which agent operation logs can be queried. Agent policy and management: a system service queries the latest Policy at install or update time, periodically or on demand, and supplies the agent operation policy. Pre-execution check: before starting an operation the agent queries the policy; a system service verifies the agent’s identity and returns the final policy. Agent log management: decision and operation logs go to a log service for storage and management, and are available back to the app.

SAEP provides three main scopes for rules: App (rules designed at the level of the whole application), Activity (rules for a specific page, such as browsing products; essentially the page or interaction-interface unit of an Android app), and Intent (rules set for a specific business intent, for example allowing editing but restricting publication).

Figure 2 | The scopes of an SAEP policy
Table in Chinese with three rows. App scope matches the whole application. Activity scope matches a specified page, for example a login page that prohibits screenshots and input, or a payment confirmation page that prohibits operation entirely. Intent scope matches a specified business intent, for example restricting content modification, final publication, deletion or the claiming of incentive benefits. A paragraph below states three principles for merging policies.

Table from the original post (source: Doubao's SAEP integration documentation), in Chinese. DCC's English reading follows.

DCC’s English reading. App policy can be expressed in three scopes.

ScopeMatchesExample
AppThe whole applicationThe agent may perform the operation on any page of this app
ActivityA specified pageThe login page prohibits screenshots and input; the payment confirmation page prohibits operation entirely
IntentA specified business intentRestrict content modification, final publication, deletion, or the claiming of incentive benefits

Policy merging follows three stable principles: the system baseline cannot be relaxed; where several rules match, the stricter value applies; and on a parsing failure the system falls back to an available default policy. Because the scheme’s own materials differ in how they describe the ordering of App, Activity and Intent, developers should not rely on “later entries override earlier ones”. Sensitive pages should declare their restrictions explicitly, and the final Effective Policy should be confirmed with acceptance test cases on the target image.

The integration documentation lists seven restriction fields, which fall broadly into two classes. One class is aimed at technical capabilities, such as reading the screen and simulating taps and input, and is closely tied to operating-system interfaces. The other is aimed at business conduct, including publishing, deleting and claiming rewards; here the system has to judge which class of business intent the agent’s current action belongs to, which involves more complex semantic recognition. The documentation also notes that whether intent-level constraints already have complete system enforcement depends on the target version, and that acceptance cannot be completed merely on the agent’s reply of “complied”. A reminder: the English field names have to be read in the context of agent operation, or they are ambiguous. The integration documentation explains the typical use of each.

Figure 3 | The restriction fields SAEP supports
Table in Chinese, cropped at the right edge in the original, listing seven fields, the meaning of the value true, and a typical use: global_disable, screenshot_disable, input_disable, modify_content, post_content, delete_content and account_incentive. A note below explains that true means restrict or prohibit and false means not restricted by that field.

Table from the original post (source: Doubao's SAEP integration documentation, § 4.3.2), in Chinese; the right-hand column is cropped in the original image. DCC's English reading follows.

DCC’s English reading.

FieldMeaning of trueTypical use
global_disableProhibit all agent operationThe whole app does not allow GUI agent operation
screenshot_disableProhibit screenshots and screen framesLogin, payment confirmation, display of private information and other sensitive pages
input_disableProhibit simulated input such as taps and text entryPassword, verification code and permission-change pages
modify_contentRestrict modifying content or settingsNickname, contacts, permissions, device and payment settings
post_contentRestrict final sending or publishingThe agent may help edit but may not complete the sending for the user
delete_contentRestrict final deletionThe agent may help locate or select the item but may not delete it
account_incentiveRestrict operations on incentive benefitsCheck-ins, prize draws, flash sales, tasks, points

Uniform semantics: for every field, true means “restrict or prohibit” and false means “not restricted by this field”; an unset field is treated as false. false does not mean the system will in the end allow the operation: the system security baseline and other stricter rules may still refuse it, as in payment scenarios. (The last words of several “typical use” cells are cut off in the source image; DCC has completed them from context.)

The system combines the agent’s identity and trust status, the security baseline, the app’s declaration and the user’s authorization to reach a decision of ALLOW, BLOCK, CALL_USER and so on, then applies the constraint at the entry points to key system capabilities and generates an audit record.

Figure 4 | The system decisions SAEP defines
Table in Chinese under the heading Decision, enforcement and user takeover. Four decisions are listed with the system behaviour and what the app and the agent should expect: ALLOW, BLOCK, CALL_USER and CLARIFY. An introductory paragraph says enforcement should sit on critical paths controlled by the system.

Table from the original post (source: Doubao's SAEP integration documentation, § 2.5), in Chinese. DCC's English reading follows.

DCC’s English reading. Decision, enforcement and user takeover. Enforcement should sit on critical paths controlled by the system. For example, app-level disabling can prevent the target app from entering the agent’s virtual screen; a page-level screenshot prohibition can add a secure layer; input restrictions can block simulated injection on paths such as InputManager. Whether intent-level constraints already have complete system enforcement depends on the target version, and acceptance cannot be completed merely on the agent’s reply of “complied”.

DecisionSystem behaviourWhat the app and the agent should expect
ALLOWAllow the corresponding system capability to be called, and record the normal execution resultThe operation must still pass the app’s own business authentication; it does not dispense with user authorization
BLOCKBlock the operation at the policy enforcement point (PEP), return a clear refusal, and do not fall back to a bypass capabilityStop the automated operation, explain the restriction to the user, and do not try another channel to get around it
CALL_USERPause automation and show a system-level confirmation or takeover interfaceWait for the user to confirm or operate in person; the confirmation covers only the specific step shown
CLARIFYIn the current materials, used to express clarification or a soft constraintThe exact semantics, and whether it is open to third parties, depend on the target SDK or API package

According to the developer integration documentation Doubao has published, SAEP allows a third-party app to express access restrictions through its Manifest or through dynamic policy, and to set boundaries at the three levels of App, Activity and Intent. This already has the outline of a basic access-control architecture, an attempt to move automated access toward permission management that is machine-readable and enforceable. At the same time, SAEP concentrates several key discretions in the system control layer. To judge whether SAEP is enough to win the trust of third-party apps, one should therefore start from how access-control power is allocated.

First, under the integration documentation an app may choose a global prohibition or restrict specific conduct, but complex business logic has to be translated into the protocol’s preset fields. If the system cannot adequately recognize the multiple states to which a given action may map semantically, the app may choose coarser restrictions to avoid risk. The official documentation also warns that some nested schemas, priorities and Intent-level enforcement capabilities still have to be checked against the target version.

Second, a GUI agent submits its target, action and context before it operates, but there is no stable one-to-one correspondence between a technical action and a business intent. If the system relies too heavily on the agent’s own account, there is a risk of misclassification. The authoritative semantics of a key operation ought to be supplied by the app, with the system then deciding in light of the user’s mandate and the security baseline.

Third, where an agent’s conduct is not trusted, SAEP tries to block it on system paths such as screenshots, accessibility reads, taps and input injection. But what is the remedy when the protocol is unavailable, the system is upgraded, policy parsing fails or the cache has not been updated? The Doubao phone assistant’s official documentation currently requires apps to detect the protocol’s status and keep their own security controls, so the concern about continuity of enforcement remains.

Finally, the audit log records the agent’s identity, the policy version, the ruling and the execution result, and is an important basis for accountability. But where the system takes part in ruling, enforcement and the generation of evidence all at once, only event identifiers that can be verified across parties, trusted timestamps, proof of log integrity and a route for complaint and correction can take it beyond one party vouching for itself, and produce real trust.

SAEP offers an agent access-control scheme of real substance. Through app policy declaration, graded restriction, system ruling and enforcement, a third-party app can convert its own operating boundary into rules a machine can recognize, and an agent’s cross-app access is thereby given fairly clear technical constraints. Compared with relying purely on the agent’s self-discipline, or with each app adopting its own blocking measures, the attempt has positive value. It also supplies a technical basis for agent authorization to move from unilateral promise toward systematic governance.

Whether SAEP can effectively ease third-party apps’ security concerns, however, depends on several conditions still to be verified. First, whether the protocol’s preset rule fields can adequately express complex business restrictions, especially the correspondence between technical actions and business intents. Second, whether the system can accurately understand, promptly update and continuously enforce app policy, so as to prevent rules from lapsing, recognition errors or omissions in enforcement. Third, whether policy rulings, execution results and the related evidence can be independently verified across parties, so that app operators need not rely entirely on the protocol implementer’s account of its own conduct. Finally, whether different agents receive equal treatment in identity authentication, the application of policy and the invocation of system capabilities also awaits an objective standard of testing.

As an access-control mechanism, what SAEP has not yet solved is whether policy expression, decision and enforcement are reliable from the standpoint of other market players. As a cross-ecosystem interoperability arrangement, its further development must also solve the problem of institutional trust raised by the relative concentration of the powers to make, interpret and enforce the rules. The former can be improved gradually through technical iteration. The latter depends on multi-party participation, common standards and independent oversight.

Even if SAEP were fully optimized at the technical level, third-party apps might still choose to refuse out of considerations of market competition, control of the entry point and the distribution of benefits. Technical trust is therefore an important condition for interoperability, but it is not enough to ensure that cross-ecosystem cooperation happens naturally. To understand further why refusal is now so widespread, the view must move from the protocol’s architecture to the competitive relationship between agents and the existing app ecosystem, and to the incentives for cooperation.

Beyond technical trust: the co-opetition dilemma of interoperability

Setting technical and architectural detail and the trust it earns to one side, this part takes up a macro question: would the market have reacted this way however the architecture was designed? It must be stressed first that this is not only a game between AI vendors and traditional apps as new replaces old. Today’s large digital platforms often own models, agent products and app ecosystems all at once. Their endowments differ: social relationships, content distribution, e-commerce transactions, local services and payment capability are each someone’s strength. Second, choosing to refuse does not mean not wanting to take part. Mostly it follows the logic of wanting a bigger share of the cake for oneself.

The dispute over agent interoperability that SAEP has set off shows, to some degree, the features of the Prisoner’s Dilemma of classical game theory. For the large internet platforms, cross-ecosystem agent interoperability promises to enlarge the market for digital services, lower users’ costs of operating across platforms and create new transaction opportunities. Yet as agents gradually become the main entry point through which users reach digital services, opening one’s own app ecosystem may also mean ceding part of one’s control over the allocation of traffic, over user relationships and over transactions. Each party may therefore hope that the other platforms will open to its agent while keeping its own app ecosystem closed. If all choose this way, the result may be an inefficient equilibrium of mutual closure. Each has kept its existing boundary, and together they have lost the incremental gains that cross-ecosystem interoperability could create.

[DCC: condensed. The author recounts the model’s origin in experiments by Melvin Dresher and Merrill Flood at the RAND Corporation in 1950 and Albert W. Tucker’s retelling of it as the story of two prisoners. Its lesson: even where all parties see the common gain from cooperation, individually rational choices made without effective constraint may still produce a collectively inefficient result.]

To classify the situation directly as a Prisoner’s Dilemma, several strict premises still have to be met. First, the participants should have relatively symmetric market positions and payoff structures; at the least they must share the same ordering of strategic preferences. Second, staying closed must yield a higher individual payoff whether or not the competitor opens; that is, closure must be each party’s dominant strategy. Third, the payoff created when both open should be higher than when both stay closed. To isolate this incentive structure, one must also set aside legitimate limiting factors such as app security, user privacy, technical adaptation and the allocation of liability, as well as the effect on strategic choice of long-term cooperation, reciprocal commitments and outside regulation. In reality the platforms’ endowments, business structures and agent capabilities differ widely, and so do the costs each bears and the gains each may obtain from opening its apps. Some platforms may gain independently through new user reach and transaction channels even if they open first. The Prisoner’s Dilemma thus shows that unilateral rationality may lead to collective inefficiency, but it is not enough to explain fully the different choices made in the SAEP dispute.

The mass refusal of the SAEP protocol also has some of the features of the Stag Hunt.

[DCC: condensed. The author traces the Stag Hunt to Rousseau’s 1755 Discourse on the Origin and Basis of Inequality Among Men and to Brian Skyrms’s 2004 book The Stag Hunt and the Evolution of Social Structure. The hunters gain most by holding their posts and taking the stag together, but each may abandon the hunt for a hare that passes within reach.]

What the Stag Hunt reveals is the tension between the gains from cooperation and the risks of cooperation. All parties know that cooperating together brings the richer reward, but because none can be sure that the others will keep their commitment, each turns to independent action with a lower but safer payoff. In agent interoperability, an app platform may not oppose opening as such. It may worry that after it opens first, its competitors will stay closed, or that it will have to bear alone the costs of technical security, compliance liability and adjustment of its ecosystem. Through app policy declaration, agent identity authentication, system enforcement, user takeover and operation audit, SAEP tries to reduce the uncertainty of cross-ecosystem access, and in a sense it is responding to exactly this coordination problem. For market players that are willing to interoperate but lack sufficient trust, common technical standards, credible commitments and risk-sharing mechanisms may be important conditions for cooperation.

The Stag Hunt has its own premises, however. First, the parties must accept that mutual opening is better than mutual closure, and be willing to open when sure that the other side will open. Second, the gains from opening unilaterally must not be enough to offset the costs and risks borne alone, so that a participant’s best choice depends on whether the others cooperate. Third, one must set aside for the moment, or weaken substantially, the gain from unilateral closure that competition over the agent entry point brings. If a platform would still prefer to restrict a competitor’s agent from its ecosystem even when sure the competitor will open, the problem has gone beyond coordination and trust alone. Conversely, agent interoperability does not necessarily need all platforms to open at once in order to create value. For some apps, being early to connect to a new agent entry point may itself bring extra traffic and transaction revenue. These realities limit how much the Stag Hunt can explain.

The fundamental difference between the Prisoner’s Dilemma and the Stag Hunt lies in the participants’ judgment of the gains from cooperation and the strategy that follows. In the Prisoner’s Dilemma, even when sure the other side will cooperate, the individual still has an incentive to defect. In the Stag Hunt, so long as the individual can be sure the other side will cooperate, it is willing to cooperate. The former is about incentive structure and the distribution of benefits; the latter is about coordination failure and insufficient trust. This difference also determines the direction of the institutional and regulatory response. For a Stag Hunt, transparent rules, common standards, credible certification and risk-sharing help to bring about cooperation. For a Prisoner’s Dilemma, building trust or agreeing an initiative is not enough to change the individual’s best choice; the payoff structure of each party must also be adjusted through reciprocal arrangements, non-discrimination obligations and the necessary external constraints.

In essence, competition in the agent market has two components. On one side, every model vendor hopes the agent market will grow and create new revenue, but fears bearing too much risk by opening first. On the other, although relatively independent model vendors exist, the main technology oligarchs each want to be the principal agent entry point, and even if the whole AI ecosystem could create greater gains, they have a motive to restrict competitors’ agents. The former has the features of a Stag Hunt: both sides believe cross-ecosystem interoperability can raise the service capability of their own agents and create new user demand and transaction opportunities. The latter has the features of a Prisoner’s Dilemma: both sides want their own agent to be the main entry point through which users reach digital services of every kind. For one and the same platform, therefore, opening has two opposite effects: market expansion (more apps can be invoked, and the agent’s improved service capability promotes transactions and user demand) and entry-point substitution (user interaction shifts to another agent, and the transfer of control over traffic and data lowers the “rent”).

[DCC: condensed. The author introduces asymmetric games, citing John Maynard Smith and Geoffrey Parker’s 1976 paper “The Logic of Asymmetric Contests” and Maynard Smith’s 1982 Evolution and the Theory of Games. The point taken from them: facing the same opportunity to cooperate, different players may form very different strategic preferences because their existing resources, capacity to act and share of the payoff differ.]

Bringing this view to the agent interoperability market, the SAEP dispute may also present a more complex asymmetric game. Today’s large digital platforms often own models, agent products and app ecosystems at the same time, with different strengths in social relationships, content distribution, e-commerce, local services and payments. Any party’s agent, to complete a complex task, may depend on app resources controlled by a competitor, while letting an outside agent into one’s own apps may change the user entry point, data relationships and the distribution of transactions. Two gains pointing in opposite directions result. Interoperability can enlarge service capability and add transaction opportunities and user convenience, but it may also weaken the existing app’s position as the entry point for user interaction. For a given platform, the most favourable situation may be one in which its own agent can enter other ecosystems while competitors find it hard to enter its own. Mutual dependence thus coexists with competition over the entry point, and the conditions of opening become the subject of negotiation.

For a company whose main product is an agent service, the opening of third-party apps directly enlarges the agent’s range of service and raises the product’s appeal. For a platform that already holds a huge user base, a closed transaction loop and a traffic entry point, allowing outside agents access may weaken its existing control over user interaction and the distribution of value. Even among large internet companies, agent capability, app resources and the degree of cross-ecosystem dependence differ markedly. One player may therefore always lean toward opening and another always toward closure, and their strategic choices need not form a symmetric structure of cooperation or defection. The choices apps have actually made offer evidence for this. 155 apps expressly refused automated operation by the Doubao phone assistant, but not every large platform took the same position. JD.com, Ctrip, Tongcheng and others were not on the first refusal list, and some operations have been carried out in media tests, while the different apps under Baidu have also taken different attitudes.

The Stag Hunt helps to explain the coordination risk: each side believes mutual opening pays more, but fears opening first without the other following. The Prisoner’s Dilemma points to another possibility: even if the other side opens, one may still lean toward closure in order to keep the entry-point advantage. But both classical models presuppose a particular payoff structure and cannot fully explain the present standoff over agent authorization. Between platforms there are differentiated resources and asymmetric payoffs, and also a range of strategies such as graded authorization, selective opening and conditional access. Game theory offers useful tools for understanding the dynamics of today’s agent market and the possible reasons why so many companies refuse to interoperate, but its explanatory power depends on an accurate grasp of the real payoff structure. At least three interrelated variables have to be examined.

First, a platform’s control over its own app ecosystem. This includes the user scale, key app resources, closed transaction loop, data and distribution channels the platform holds, and how substitutable the relevant services are for other ecosystems. The stronger the control, the greater the existing interests the platform may keep by staying closed, and the stronger its bargaining power in refusing outside agents may be. Rich app resources do not, however, automatically mean closure pays more; the new market opportunities that opening may bring still have to be examined.

Second, a platform’s competitiveness as an agent entry point. This involves its foundation model, agent product, capability to operate on the terminal, user base and capability to execute cross-app tasks. A platform with strong agent entry-point competitiveness may wish at once to widen its own agent’s access and to restrict competing agents from its app ecosystem. Conversely, an app service provider without an independent agent entry-point advantage may be more willing to use an outside agent to gain new user reach and transaction opportunities. Different combinations of agent capability and control over an app ecosystem will affect each party’s judgment of the payoffs from opening and from closure.

Third, the degree of cross-ecosystem complementarity. This variable asks whether the app resources, service capabilities and user needs of different players can form real complementarity, and whether the incremental gains created by interoperability can be reasonably shared among the participants. Where an agent can bring an app new demand, transactions and service opportunities, opening may have independent commercial value. Where interoperability mainly takes the form of a transfer of the user entry point, and the app service provider can hardly share in the resulting gains, the incentive to stay closed may strengthen. Existing commercial cooperation, technical adaptation and liability-sharing arrangements may also affect whether complementary gains are actually realized.

These three variables interact, so that the agent interoperability market shows a pattern of interwoven cooperation and competition and a highly heterogeneous payoff structure. The form of the game may vary by company, by application scenario and even by the specific type of operation. Even one and the same platform may make different judgments about opening for different operations such as information queries, content generation, transaction confirmation and payment execution.

In essence, interoperability in the agent economy is a co-opetition game among different digital ecosystems over interoperability, control of the entry point and the power to make the rules. Platforms are not homogeneous, symmetric participants, and opening and closure are not a simple binary choice. Restricting access by a certain class of agent is not necessarily betrayal, and opening is not necessarily the same as cooperation. A decision taken at any one point in time may be adjusted and corrected as the market develops. For some platforms, therefore, the main obstacle may be the lack of credible commitments; for others, it may be the loss of entry-point control after opening. Still other apps may have concerns they genuinely cannot avoid, such as identity authentication, privacy protection and transaction risk control.

After SAEP was refused

SAEP was the brave step of the first to eat the crab. But this is a crab one person cannot finish, and cannot eat alone. The next question is how, and whether, the market and the regulators need to respond.

Platforms in the agent interoperability market are not homogeneous partners in cooperation. They are competitors that hold differentiated app resources, agent entry-point capability and rule-making power all at once. Because platforms differ markedly in endowment, cost of opening, outside dependence and the gains from entry-point competition, the same interoperability rule may produce different economic incentives, and may even lead to an asymmetric equilibrium in which one party leans toward opening and another toward closure. Interoperability governance therefore cannot rely only on formally reciprocal opening. I had talked this over with colleagues before, and someone immediately thought of the regulatory tools on refusal of access they had come across in competition law.

As described above, the reasons why so many market players refused are complex and multidimensional. In the end, though, they can be reduced to the market game (everyone wants to eat this cake, alone if possible, or does not want others to get to it first), and to the conservative, wait-and-see posture adopted in an ecosystem of social distrust, which has been very common over the past several years.

Specifically, three kinds of refusal can be distinguished on this basis. The first rests on definite security and compliance risks, such as account authentication, payments and transfers, bulk sending or the reading of sensitive information, and is common in the market. The second arises from structural distrust: unclear protocol rules, unverifiable enforcement, unclear allocation of liability. It has not so far been clearly voiced in public, and it is the basis on which the third part of this essay tries to build its analysis. The third uses a formal security reason to impose differential treatment on a competing agent that presents comparable risk.

The three kinds of refusal differ in their standard of proof and in how they should be governed, and the ways in which the market adjusts and intervenes differ too. Where there is no consensus on how rule semantics, identity authentication and logging capability are to be implemented, the gap can be repaired through standards, testing and industry negotiation.

Where the wording of permissions is unclear and the logic of authorization muddled, the answer can be a compliance system built on risk grading, least privilege, user takeover and traceable accountability. The most complex case is differential treatment that lacks an objective reason, or the exclusion of a competitor’s access for the sake of competition. Intervention in this problem, too, should follow the laws of the market and competition theory, and regulatory tools should be used with caution. In the negotiation stage, before any situation of competitive exclusion exists, choosing to open or to stay closed is an exercise of autonomy, and there is no justification for mandatory opening.

When I discussed agent authorization with colleagues earlier, someone immediately raised the regulatory interventions of competition law against refusal to deal, refusal of access and obstacles to technical interoperability. The association is natural, but it lacks careful thought. A digital platform can restrict competitors’ entry into adjacent markets by controlling technical interfaces, the operating system or key app resources. For an undertaking with a dominant market position, the Anti-Monopoly Law can, under specific conditions, require it to remove unreasonable obstacles to interoperability. But can third-party apps be required on that basis to accept automated operation by outside agents across the board? In principle, regulatory mandatory access may improve the conditions of competition, yet it may not raise user welfare correspondingly. Where the technical standard needed for access is proposed or led by a particular player, regulatory intervention may even entrench a new structure of market power.

In Interop, the classic work on interoperability in the digital age, John Palfrey and Urs Gasser discuss the benefits and costs of interoperability systematically. The two authors affirm the positive role of interoperability for innovation, competition, user choice and the flow of information, while stressing that interoperability is not itself the final goal of policy. Connections between different systems may increase security and privacy risks, widen the reach of system failures, and create new lock-in through technical homogenization and mutual dependence. In addition, research in competition law on adversarial standardization looks at how, when regulation requires interoperability, incumbent market players move the competitive conflict into the standard-setting process, and at how the governance of standards is affected by market power.

Traditional interoperability disputes usually turn on the sharing of technical specifications, data formats or interface capabilities, whereas a GUI agent can carry out a sequence of operations with real consequences inside a third-party app. The risks involved in letting an agent read product information are plainly different from those in letting it change account settings, send messages or complete a payment. If regulation aimed only at widening access and required app operators to accept automated operation that has not been adequately verified, users would gain more choice of agents but might also bear higher risks of privacy leakage, erroneous transactions and account insecurity.

Mandatory interoperability may also carry a more hidden cost of standardization. A common technical standard can lower the cost of coordination between systems, but it may also fix the technical path of market competition too early, so that different companies have to innovate around the same set of interfaces, operation categories and certification systems. For agent technology that is still evolving quickly, an overly uniform standard may narrow the room to explore alternative architectures. What deserves more vigilance is this: when a protocol first proposed by one company is established as the de facto industry standard, or goes on to gain regulatory recognition, that company’s influence over rule semantics, version updates, certification conditions and the manner of technical enforcement may turn into a lasting market advantage. An interoperability requirement meant to do away with closed ecosystems may instead leave market participants jointly dependent on another piece of infrastructure that is hard to replace.

Whether regulatory intervention is apt therefore has to rest on identifying the specific obstacle to competition. If the problem arises from a terminal system provider with market power favouring its own agent, regulation can focus on differential treatment in identity authentication, in the invocation of system capabilities and in policy enforcement. If the problem arises from a third-party app’s refusal of an outside agent, one must go on to examine that app’s market power, its reasons for refusing, the actual competitive effect, and whether a lower-risk alternative form of access exists. If the problem is mainly that the parties lack common standards and credible commitments, it may be more appropriate for the regulator to promote the coordination of standards, support independent testing and establish a dispute-resolution mechanism than to impose a mandatory access obligation directly.

Building consensus

Mandatory opening is unreasonable. But leaving platforms to set, on their own, conditions of admission that cannot be reviewed is equally unsuited to the healthy and orderly development of the industry. The focus of governance should therefore be on restriction reasons that can be tested and access conditions that are proportionate, so that security exceptions can be respected (that is, legitimate refusals of access) and malicious competitive closure cannot easily win exemption in the name of security. Measures of this kind are common, but whether they would themselves harm the market still requires careful legal analysis. Given the bluntness and the lag of anti-monopoly tools, however, governance should consider more agile, ex ante approaches.

SAEP has already tried to connect policy declarations at the App, Activity and Intent levels with agent identity, user authorization, system ruling and execution audit, and has explored how an app declares access restrictions and how the terminal system carries out policy ruling and enforcement. But the technical completeness of a protocol is not enough to answer why an access restriction is legitimate, how different agents obtain fair treatment, and how cross-ecosystem cooperation builds a lasting and credible institutional foundation. Even if the micro-level technical architecture were perfected and ran well, cross-ecosystem opening would still not necessarily happen of itself. Building consensus, moreover, needs more than technical rules. It needs higher-level principles that answer why a restriction is reasonable, whether similar players are treated alike, how the two sides of an access relationship share the duties of cooperation, and who has the power to review independently the conduct of the rule-maker itself.

In view of this, I think what the market needs now is to form a base-level consensus at the market level, to improve the incentives for cross-ecosystem cooperation through appropriate reciprocal arrangements, and on that foundation to let reasonable security and business boundaries take shape gradually through negotiation and market maturity.

In summary, agent interoperability needs to combine reasonable procedural governance, reasonable risk conditions and risk perception (grading), non-discriminatory access standards, neutral standard formation and reciprocal ecosystem arrangements, and to build independent verifiability into the means of implementation, so as finally to improve the incentives of different players to cooperate.

Specifically:

First, establish minimum general interoperability rules: on the basis of SAEP, make clear the identity authentication, user authorization, security baseline, log audit and complaint mechanisms, while watching and measuring the allocation of power in real scenarios.

Second, allow differentiated conditions to be set for apps and operations of different risk levels, and require those conditions to be objective, transparent and verifiable.

Third, explore reciprocity mechanisms, without pursuing complete equivalence of resources or commercial value.

Fourth, for platforms that hold significant market power and control key app resources, study whether their refusal to interoperate constitutes an unreasonable restriction of competition, but do not impose the same mandatory opening obligation on all platforms across the board.

Fifth, establish a multi-party standards-governance mechanism, to prevent the maker of an interoperability protocol from using its rule advantage to exclude competitors.

On this basis, I am working to bring industry, academia and research together around a consensus declaration, so that the industry’s development rests on a set of minimum basic rules that all can accept, takes one small step further, and avoids vicious competition. The declaration is to take reasonableness, non-discrimination and functional reciprocity as its core, and to promote common rules that take account of user authorization, app rights and interests, technical security and market competition. Its point is to give different technical protocols an evaluation standard they can all accept, so that the degree of interoperability and the way it is achieved can keep adjusting with technical development, the understanding of risk and market conditions.

The order of interoperability in the age of agents cannot, in the end, be completed by any one company’s protocol alone. It needs, while respecting lawful boundaries, to make the conditions of opening foreseeable, the reasons for restriction testable, the gains and risks of cooperation shareable, and disputes correctable. I hope this declaration can be a practical starting point for building the ecosystem.


DCC’s reading

The notes below are DCC’s, not the author’s.

  • MIIT has since set the default the other way. SAEP treated silence after 30 days as permission. Part I, item (4) of 工信厅信管〔2026〕42号 requires an app developer’s declared consent before an agent may invoke the app. Li’s five proposals assume the default is still open to negotiation; on the question of who must speak first, it no longer is. See DCC’s map of the Notice.
  • His three kinds of refusal still matter. The Notice asks only whether consent has been declared. It gives the app developer no duty to state reasons and gives the agent no route to challenge a refusal. The third kind of refusal, security used as a pretext against a competitor, is left to competition law: the Anti-Monopoly Law, which Li discusses, and the Anti-Unfair Competition Law, under which China’s platform-data cases have so far been decided.
  • The protocol detail is useful on its own terms. The App, Activity and Intent scopes, the seven restriction fields and the four decisions are the most concrete public description DCC has seen of how an app’s refusal can be expressed to an on-device agent. Any declaration mechanism built under the Notice will need something like them. DCC has not reviewed Doubao’s developer documentation directly; the description is the author’s.

— 李汶龙 (Li Wenlong), 智能体授权的博弈困境破局——从SAEP到合理、非歧视、 互惠基础互操作规则 (Breaking the Game-Theoretic Deadlock of Agent Authorization: From SAEP to Reasonable, Non-Discriminatory and Reciprocal Basic Interoperability Rules), 科技利维坦, October 26, 2026. Translated by DCC in full except for three condensed background passages, marked in the text; decorative images, two book covers and the list of the author’s earlier posts are omitted. Original (Chinese).

Market facts, press reports and the description of the SAEP documentation are as given by the original author; DCC has not independently verified them.

— Not legal advice.

— Not legal advice.


§ SUBSCRIBE

The Monday brief.

One short email every Monday. New briefs on Chinese data-compliance rules from the previous week, with the source law cited.

Opt-in only. Unsubscribe anytime by replying "unsubscribe" to any issue.

SUPPORT DCC

Keep the publication free to read. Suggested support is $19.99, or choose your own amount.

Support →