The Journey to Agentic Payment: The Real Challenge Lies in the Middle Path

As the industry discusses the endgame of Agentic Payment, the real question worth studying is not what the future looks like, but how the path from "human confirmation" to "machine-executed payment" will be paved step by step. Payment has never been about "can it be done," but rather "can it be authorized, constrained, and trusted."

Written by: Yuki

Recently, there has been a surge of developments surrounding the infrastructure of Agentic Payment.

Tempo's mainnet has launched, Stripe has partnered with Tempo to introduce the Machine Payments Protocol (MPP); Visa is advancing the Visa Intelligent Commerce and Trusted Agent Protocol; Google has launched the Agent Payments Protocol (AP2) and continues to extend towards the Universal Commerce Protocol (UCP); Coinbase has also released Agentic Wallets and x402, attempting to provide native wallets and payment capabilities for AI agents. In other words, payment protocols, identity trust layers, commerce standards, and agent wallet layers around Agentic Payment are beginning to emerge simultaneously.

As these infrastructures begin to appear, the industry's descriptions of the endgame are also becoming increasingly aligned.

In the future, AI agents will not only assist people in searching, recommending, and placing orders but will also directly complete transactions, payments, and settlements on behalf of users or businesses within authorized boundaries; the initiators of payments will gradually shift from "human manual confirmation" to "machine execution according to rules." Recent public statements from Visa, PayPal, and OpenAI regarding agentic commerce are essentially converging in this direction.

Everyone is now discussing the endgame quite similarly. However, the biggest uncertainty now is not whether the endgame exists, but which layer will be validated first. Even platforms as close to traffic entry as OpenAI are still adjusting the implementation of "direct checkout," indicating that while industry consensus has formed, the execution path is far from finalized.

The real question worth studying is no longer what the future will look like, but: How will we pave the path from today to that future?

Because payment is not search, recommendation, or order placement. The real difficulty in payment has never been about "can it be initiated," but rather "can it be authorized, constrained, audited, and held accountable."

Therefore, this article does not wish to reiterate the endgame of Agentic Payment. Instead, it aims to dissect: how will the path from "human confirmation" to "machine automatic payment" unfold, which layers will emerge first, and where will it encounter key bottlenecks?

1. Agentic Payment Will Not Happen All at Once; It Will Unfold in Phases

---

Many people discussing Agentic Payment assume a premise: as long as the model is smart enough and the payment interface is open enough, this will naturally happen.

However, payment has never been a simple capability issue, but rather a trust issue. AI can certainly quickly learn to search, compare, place orders, and call interfaces, but that does not mean the system will allow it to directly control the flow of funds. Payment involves not just the action itself but also authorization, responsibility, risk control, compliance, and settlement. For this reason, Agentic Payment cannot suddenly enter a state of "fully machine-executed payments"; it is more likely to unfold along a path of gradually releasing permissions.

In other words, the real question is not "when will AI start making payments," but rather: Under what conditions will humans gradually transfer payment authority to machines? If we break down this path, it will roughly go through five stages.

Stage 1: AI Participates in Decision-Making but Does Not Touch Payments

This is the earliest and easiest stage to accept. AI agents first enter the pre-transaction chain, helping users complete searches, filtering, recommending, comparing prices, and preparing to place orders, but the final payment step is still confirmed by humans.

Today, many public demonstrations of agentic commerce essentially remain at this stage: AI can push the purchasing action very close to completion, but the actual payment confirmation is still in human hands. PayPal, when explaining agentic commerce to consumers, also clearly states that "users can request to confirm again before purchasing" as a standard form.

Why does this step happen first? The reason is simple: it has the least impact on the existing payment system. Merchants do not need to completely overhaul their systems, and payment networks do not need to immediately rewrite authorization logic. AI merely upgrades from a "recommendation tool" to a "transaction orchestration tool."

But its boundaries are also very clear: AI can help you get to the payment point, but it has not yet truly obtained payment authority.

Stage 2: AI Can Trigger Transaction Processes, but Humans Retain Final Confirmation Authority

Taking a step further, AI is not just giving suggestions but can directly advance the transaction process. It can create orders on your behalf, call merchant interfaces, prepare payment parameters, and even push the purchasing action to the final confirmation step.

This will be a very long transitional phase, as it balances efficiency and a sense of security. Users begin to accept "AI helping me complete the process," but have not yet accepted "AI deciding to spend money on its own."

Visa's recent actions surrounding Intelligent Commerce and Trusted Agent Protocol are essentially paving the way for this transitional state: on one hand, enabling merchants to identify trustworthy AI agents, and on the other hand, still basing transactions on controlled and verifiable authorization mechanisms. Visa defines the Trusted Agent Protocol as a core component of Intelligent Commerce, focusing on enabling merchants to identify and verify "trustworthy agents with genuine transaction intentions."

Why is this step important? Because it begins to transform payment from "manual human operation" to "machine orchestrated processes." However, it still has not completed the most critical leap: machines have not yet obtained continuous payment authority.

Stage 3: AI Receives Authorized Limits Within Defined Boundaries

This is the first real turning point. In this stage, humans no longer confirm each transaction individually but first define rules and then allow machines to execute within those rules. Authorization begins to shift from "action authorization" to "policy authorization."

For example, users or businesses can stipulate: you can only make payments within the monthly budget; only pay to whitelisted merchants; only purchase specific categories; must confirm again for amounts exceeding a certain threshold; certain payments can only occur within designated time windows.

At this point, AI does not receive "free payment authority" but rather a constrained execution authority. The core of Agentic Payment also shifts for the first time from "intelligent assistant" to "machine agent."

When Google launches AP2 in 2025, it clearly defines it as an "open protocol for securely initiating and executing agent-led payments across platforms"; by January 2026, Google continues to launch UCP, positioning it as an open standard for agentic commerce and emphasizing compatibility with AP2. This evolution itself indicates that the industry is beginning to understand "AI payments" as a complete set of authority and process systems, rather than a single payment API.

Therefore, what is truly important in this stage is not who first creates an AI payment button, but who first turns the machine authorization system into a product.

Stage 4: AI Has Native Callable Wallets or Payment Accounts

At this level, things begin to change significantly. In the previous stages, AI was essentially still "using human payment tools on behalf of humans." But at this stage, AI no longer just calls a card, a wallet balance, or a checkout entry, but begins to possess more native financial containers. It can hold, receive, allocate, and spend funds under clear rules, upgrading payment capability from a checkout action to a continuous financial management capability.

When Coinbase launches Agentic Wallets in February 2026, the definition provided is very straightforward: this is a wallet infrastructure built specifically for AI agents, supporting autonomous spending, earning, and trading, with x402 positioned at the core of this system as a payment protocol for autonomous AI use cases. The official statement also emphasizes that x402 has been validated in scenarios such as machine-to-machine payments, API paywalls, and programmatic resource access.

The significance of this step is that the economic behavior of agents is no longer just "completing a purchase on behalf of someone" but begins to possess continuous income and expenditure and resource scheduling capabilities. Payment thus transforms from a front-end consumption behavior into an account capability within the machine economy.

Stage 5: Continuous Machine-to-Machine Transactions Become the New Norm

This is the endgame that everyone loves to discuss today, and it is also the easiest to imagine but the hardest to fully implement at once. In this stage, both parties in the transaction may no longer be traditional "humans." It could be one agent paying API fees to another agent, a corporate system automatically replenishing cloud resource quotas, or real-time settlements occurring when models call models, devices call services, or workflows call data sources.

The characteristics of such payments are clear: small, high-frequency, automated, cross-system, 24/7, and often inherently cross-border. Because of this, the requirements for underlying infrastructure differ from traditional human payments.

The launch of Tempo's mainnet and the introduction of MPP are centered around establishing a foundation: Tempo clearly positions itself as a payments-focused L1, while MPP is defined as an open standard for machine payments, supporting payment methods like stablecoins and cards; Stripe's official documentation directly describes MPP as "an internet payment protocol for machine-to-machine payments."

However, although the endgame has been clearly articulated, it does not mean that it has fully materialized today. A more realistic assessment is that the industry is advancing from the first and second phases towards the third and fourth phases, while the fifth phase still resembles a gradually approaching direction. Coindesk's report on x402 in March 2026 even pointed out a stark reality: technology is advancing, but real demand and large-scale usage have not fully caught up.

Before diving into the specifics of the players, it is helpful to take a more comprehensive layered look at the infrastructure of this round of Agentic Payment. You will find that while it appears bustling, it is not all working on the same thing, but rather advancing simultaneously across different layers.

Agentic Payment is not a product, but a whole set of infrastructure stacks that are growing simultaneously. From protocols, identities, and standards to wallets, settlements, and currency layers, the key pieces of Agentic Payment have already begun to take shape in the hands of different players.

The Agentic Payment Table is Set, but Each Player is Betting on Different Cards

---

After breaking down these five phases, looking back at the market players makes it easier to understand why everyone is talking about Agentic Payment, yet their actions seem different. They are not betting on the same thing.

Stripe / Tempo: Betting on "Machine Payment Protocol + New Settlement Foundation"

On the Stripe and Tempo side, the most noteworthy aspect is not a single product action, but their attempt to occupy two layers simultaneously: the upper layer of machine payments protocol and the lower layer of machine-native settlement rail.

MPP is the protocol layer, attempting to define how agents and services initiate, accept, and coordinate automated payments; Tempo is the foundation layer, aiming to meet the cost, confirmation speed, and scalability requirements for such payments. After the launch of Tempo's mainnet, the official statement has been very clear: if stablecoins become the core layer of internet commerce, the infrastructure for mobile currency needs to be specifically built for payments rather than general on-chain computation.

In other words, Stripe / Tempo is not just trying to create "payment functions usable by AI"; they are betting that the machine economy will require a more native payment protocol and settlement layer.

Visa: Betting on "Trusted Agent Identity + Merchant-side Access Order"

Visa's approach differs from that of Stripe / Tempo. It does not start with new chains or new wallets; instead, it addresses a more immediate issue: as more AI agents begin to access merchants and complete purchases on behalf of users, how can merchants distinguish between "trusted agents" and "malicious bots"?

The Trusted Agent Protocol and Visa Intelligent Commerce essentially answer this question. Visa aims to occupy the most critical yet often overlooked layer in agentic commerce: the trust layer and merchant access order. From Visa's public statements, it is clear that they are not simply placing an AI payment button; they want to maintain a recognizable, verifiable, and accountable transaction relationship among merchants, issuers, and consumers. Visa seems to be focusing on the second and third phases: first making agents "trusted transaction participants," and then gradually entering controlled authorization.

Google: Betting on "Open Commerce Standards + Cross-Protocol Coordination"

Google's position is somewhat unique. When looking at AP2 and UCP together, what it truly aims to do is not single-point payments, but rather the open language layer of agentic commerce. AP2 addresses how agent-led payments can be securely initiated, while UCP resolves how consumer surfaces, businesses, and payment providers can share a set of common commerce primitives. Google's official definition of UCP is straightforward: it is an open-source standard for the next generation of agentic commerce, compatible with AP2 and designed to work in conjunction with existing retail infrastructure.

This means Google is betting on middle-layer standards: it may not personally handle every payment, but it hopes to define the public grammar of "how agents understand products, merchants, payment capabilities, and transaction processes."

From a pathway perspective, Google seems to be making strides near the third phase: it is helping the industry standardize "machine authorization" and "cross-platform transaction coordination."

Coinbase: Betting on "Agent Native Wallet Layer"

Coinbase does not start with merchant access or open commerce protocols; instead, it directly tackles the funding container of agents.

The logic of Agentic Wallets is clear: if there will be more and more machine-native transactions in the future, then agents themselves need a programmable, constrained, spendable, and receivable native wallet layer. x402 plays the role of payment protocol here, while the wallet provides account capabilities. Coinbase officially describes this product as enabling agents to gain the ability to "spend, earn, and trade."

Thus, Coinbase is actually betting on the fourth and fifth phases: not first enabling AI to better access human payments, but first allowing AI to have its own executable account system.

Circle / Tether: They Haven't Created Protocols, but are Defining "What Money Machines Use"

If we break down this round of players further, we will find an easily overlooked position: Circle and Tether. They have not directly launched the Agentic Payment protocol, nor have they created a public chain specifically for AI agents, but they are continuously reinforcing a more foundational capability—allowing money itself to be programmatically invoked.

In recent external statements, Circle has repeatedly emphasized programmable money and API-based payments. In March 2026, Circle launched the Nanopayments testnet, directly describing it as one of the core primitives of "agentic economic activity," supporting gas-free USDC transfers as low as $0.000001.

Essentially, this addresses a question: when the initiators of transactions begin to shift from "humans" to "machines," does money itself also need to become a callable infrastructure?

If we say:

• Stripe is defining "how to pay" (payment protocol)
• Visa is defining "who can pay" (identity / trust)
• Google is defining "how transactions are understood" (commerce standard)
• Coinbase is defining "who holds the account" (wallet layer)

Then Circle and Tether are doing something else: defining what money machines use to pay. The role of stablecoins here is actually closer to "base currency" than "payment tool."

Because only when money itself can be called via APIs, can be embedded in programs, and can flow across systems at low cost, does machine payment have a realistic foundation. In other words, Agentic Payment largely relies on the premise that "money has already been digitized and is programmable."

However, there is also an easily overlooked boundary: making money programmable does not mean machines can safely spend money. True Agentic Payment still requires:

• Authorization systems
• Identity verification
• Risk control mechanisms
• Accountability

These capabilities are precisely the layers that Stripe, Visa, and Google are supplementing. Therefore, from a more complete perspective, Agentic Payment is not a capability that a single player can accomplish alone, but rather a system that is being pieced together: protocols are growing, identities are developing, wallets are emerging, and currencies are actually ready.

The real uncertainty lies not in whether there is money, but in when this system will be truly pieced together and begin to support real transactions.

PayPal / OpenAI: Closer to Transaction Entry and Front-End Interaction Layers

PayPal's explanation of agentic commerce leans more towards the actual transaction experience in the consumer and merchant world, emphasizing how AI can assist users in completing purchases and how varying degrees of authorization models can unfold in checkout scenarios. On the OpenAI side, it is closer to the front-end entry: turning conversational interfaces like ChatGPT into transaction flow entry points, and linking them with merchant/checkout systems.

But a noteworthy signal has recently emerged at this level: Google has continued to strengthen the open commerce protocol in early 2026, rather than locking all purchases into a single platform's closed loop; this itself indicates that the platforms closest to the traffic entry are still exploring whether to serve as a 'direct transaction entry' or an 'open trading orchestration entry.' This precisely highlights a problem: while a consensus on the endgame has formed, the execution path has yet to solidify.

3. Everyone is Painting Big Pictures, but Ultimately, It's About Who Lands First

Looking at the paths and players together, it becomes clear: today's competition in Agentic Payment superficially seems to be about who launches protocols first, who pushes wallets first, and who articulates the endgame first; however, what truly determines victory may not be who tells the story most completely, but who first validates the real transaction demands.

Because payment is not an industry that can be established through demonstrations alone. A protocol can go live, a standard can be published, and a wallet can be quickly developed.

With the launch of the Tempo mainnet, MPP has already identified machine-to-machine payments as a clear direction; Google has combined AP2 and UCP into a set of open standards for agentic commerce; Coinbase has also pushed Agentic Wallets and x402 towards 'AI agent native income and expenditure capabilities.' In other words, what is lacking today is not just the 'concept,' but whether there are sufficiently high-frequency, standardized, and low-friction real payment demands behind these infrastructures.

This is also why, despite the endgame becoming clearer, the execution path remains undefined: not all imagined payment scenarios will turn into real transactions.

Some scenarios are very suitable for demos but may not necessarily form a market first. Some scenarios sound very appealing, but once they enter the real business world, they will encounter issues of authorization, dispute resolution, performance, compliance, and cost structure.

What the payment industry ultimately needs to look at is never whether "this can be demonstrated," but whether "this can happen stably many times."

Therefore, what is truly worth watching next is not which company has released a new protocol or which platform has articulated a more complete vision of agentic commerce; rather, it is whether a certain type of transaction begins to occur continuously through these new infrastructures. Are there orders, are there payments, is there repeat purchasing, and is there a continuous increase in calling frequency and settlement frequency? These questions are more important than any PowerPoint slide.

From this perspective, the first to win in the future may not necessarily be the one with the most complete story, but rather the one that first finds the real demand anchor point.

Whoever first identifies high-frequency, standardized, low-dispute, and automatable scenarios is more likely to truly run a certain layer; conversely, if there are only protocols without continuous transactions, only wallets without stable payments, and only endgame narratives without real orders, then even the most complete infrastructure stack may remain in the stage of "concept first, demand lagging." Coindesk's March report on x402 has already pointed out this reality very clearly: technology is advancing, but real demand and large-scale adoption have not fully kept up.

This is also why I believe that what is most worth watching in the industry next is not who first clarifies Agentic Payment, but who first heats up a type of real transaction. Once orders and payments in a certain scenario begin to occur stably, the previously discussed protocols, identities, authorizations, wallets, and settlements will truly transition from the "concept layer" to the "infrastructure layer." Ultimately, what Agentic Payment competes on is not who can paint the endgame better, but who can first turn a segment of real demand into a continuous payment behavior.

4. The First to Run Will Not Be the Most Attractive Scenarios

If Agentic Payment is truly going to happen, in which scenarios will it first occur? Many people's first reaction might be to think of very "big" pictures: AI completing a cross-border trip for you, from booking flights and hotels to changing tickets and compensation, with automatic payments throughout; or AI directly completing cross-border procurement and supplier settlements for enterprises, even automatically managing budget pools.

These imaginations are certainly valid, and the agentic travel narrative that Visa is currently showcasing is precisely about these high-value scenarios: agents can package flights, hotels, and discounts for cardholders, and if the itinerary changes, they can automatically change tickets and complete secure payments.

However, what is most likely to run out first is probably not these "best storytelling" scenarios. The reason is simple. The payment industry does not reward the most attractive scenarios first, but rather the ones that can form closed loops most easily.

For a scenario to run out first, it typically needs to meet at least four conditions:

• Transactions must be sufficiently standardized
• Performance must be sufficiently immediate
• Dispute rates must be sufficiently low
• Authorization boundaries must be sufficiently clear

By this standard, what is most likely to be validated first may not be "AI completing my entire trip," but rather several types of data-intensive scenarios that may not seem as flashy but are easier to form real transactions.

1. API Payments, Computing Power Calls, Data Calls: The First Stop Most Like Machine-Native Payments

If Agentic Payment has a most natural starting point, it is likely API, tools, data, and computing power as digital services.

Because these types of transactions are almost inherently suitable for machines to complete: delivery is immediate, billing is clear, performance is standardized, and in most cases, there are no complex refund and after-sales issues.

Machines do not need to handle colors, sizes, subjective preferences, and reverse logistics like buying a sofa or a piece of clothing; they only need to judge: is this interface worth calling, should this service continue to be purchased, and how much does this call cost?

This is also why when Coinbase launched x402, the first emphasis was not on "AI buying things for you," but on allowing APIs, applications, and AI agents to complete stablecoin payments directly via HTTP. The typical direction publicly provided by Coinbase is precisely the direct transactions between APIs, apps, and AI agents. The key to these scenarios is not "cool," but naturally high-frequency.

A single model call, a data pull, or an automated task execution is inherently triggered by programs. If payments can also be completed natively by programs, then the entire transaction chain will truly close. Therefore, from my own judgment, what will run out first in Agentic Payment will not be "AI helping people shop," but rather "machines paying machines."

2. Digital Goods and Subscriptions: The Second Wave Most Likely to Heat Up First

The second type that is most likely to gain traction first is digital goods, software subscriptions, content services, and cross-border SaaS renewals.

The reasons are similar to the previous category: delivery is fast, disputes are few, the logic of automatic renewal already exists, and the amounts are usually easier for users or enterprises to pre-authorize compared to physical consumption.

PayPal's recent content aimed at merchants regarding agentic commerce is also increasingly leaning in this direction: it is simultaneously creating a merchant readiness checklist while emphasizing the cross-border expansion value of digital goods and subscription businesses, directly listing software, content, and subscriptions as key types suitable for cross-border delivery and recurring payments.

Why are these scenarios important? Because they will very naturally transition Agentic Payment from "one-time payments" to "continuous payments."

Once continuous payments are entered, the previously discussed layers—authorization, identity, risk control, accounts—will truly be brought in. In other words, what truly pushes Agentic Payment into the third and fourth stages may not be one-time flashy transactions, but rather a large number of ordinary, repetitive, standardized subscription-type payments.

3. Standardized Expenditures Within Enterprises: Likely to Accept Agentic Payment Earlier Than Personal Consumption

Many people think of C-end shopping when discussing Agentic Payment. However, from a practical standpoint, B-end enterprise scenarios may actually accept it earlier. The reason is not complicated: enterprises are already accustomed to budgets, approvals, supplier whitelists, and risk control rules.

For enterprises, "letting machines spend money" is not a fundamentally unacceptable principle; the key is whether the boundaries are clear enough, whether the process can be audited, and whether responsibilities can be traced back.

Google has already placed "autonomous financial workflows" prominently in its external statements about AP2, and participants like Intuit have also explicitly mentioned that the protocol can be used to deploy more automated financial workflows.

This means that what is first validated in the enterprise world may not necessarily be "agents starting their own businesses to make money," but rather some very basic things: automatic renewal of SaaS, automatic replenishment of cloud resource quotas, automatic payment of small bills to fixed suppliers, or completing standardized procurement within a clear budget pool. These things may not be as attractive, but they are very real.

And once they are successfully implemented, their reusability will be very high. So if you ask me which type of scenario is most likely to mature the authorization system, I would say it is not the shopping cart, but rather standardized expenditure management within enterprises.

4. Travel and High-Value Consumption Will Be Eye-Catching, But Will Not Be the First to Scale Up

These scenarios will certainly be showcased extensively because they are the most visually appealing and suitable for demo presentations. One of the key focuses Visa is currently showcasing is travel: agents help users combine flights, hotels, and offers, and when itineraries change, they can automatically change tickets and complete payments.

Google is also promoting agentic shopping to retailers and merchants, emphasizing that through open standards and merchant data, high-intent users can more easily complete transactions with retailers at entry points like Search. However, for these scenarios to scale up significantly, they are actually much more challenging than digital services. Because they involve:

• More complex fulfillment chains
• Higher average transaction values
• Stronger subjective preferences
• Higher probabilities of after-sales disputes
• Greater responsibilities and risk control pressures

This is also why Visa repeatedly emphasizes the issue of trust in its research: in its 2025 consumer study targeting the United States, Australia, and New Zealand, about 85% of respondents want clear control over the data that agents can access, about 43% are concerned that agents might make incorrect purchases, and about 50% worry about "decisions being made on my behalf without my involvement." This set of data actually illustrates a reality: high-value, high-controversy, and highly subjective scenarios, while they are often used to discuss the future, will not be the first to become the starting point for large-scale validation of Agentic Payment.

5. The True Watershed: Not How Big the Scenario Looks, But Whether It Can Sustain Transactions

Ultimately, for Agentic Payment to transition from narrative to reality, it relies not on a single impressive demo, but on a category of transactions occurring continuously. Are there orders? Are there recurring payments? Is there increasing frequency of calls? This is the watershed.

So my judgment is that the first scenarios to emerge for Agentic Payment will likely be those that are low-friction, can be pre-authorized, have immediate delivery, are low in controversy, and are inherently suitable for programmatic invocation. More specifically, the priority will likely be:

• First wave: API / data / computing power / digital services;
• Second wave: subscriptions / SaaS / cross-border digital goods;
• Third wave: standardized internal corporate spending;
• Only later will it reach travel, e-commerce shopping, and more complex consumer scenarios.

This also means that what is truly worth watching next is not which platform has told a bigger shopping story, but which type of real transactions begins to run continuously on these protocols, wallets, and stablecoin networks.

Once these types of transactions start to occur stably, the previously mentioned protocols, identities, authorizations, wallets, and currency layers will truly transform from "conceptual puzzles" into "infrastructure."

Five: The Real Threshold of Agentic Payment Is Not Payment Itself, But the Authorization System

Why, despite the emergence of protocols, wallets, stablecoins, and trust layers today, has Agentic Payment not fully taken off? The answer is simple. The real difficulty of this matter has never been about "how to make payments," but rather: who has the authority to make that payment.

On the surface, the payment industry appears to be handling transactions, but at its core, it has always been about authorization. Why can cards be swiped? Because there is cardholder authorization behind it. Why can corporate procurement make payments? Because there is approval authority and budget authority behind it. Why can automatic renewals exist? Because users have given revocable authorization in advance.

Therefore, what Agentic Payment truly needs to solve is not to let AI learn to invoke payment interfaces, but to allow the financial system to accept a new question:

When the initiator of a transaction becomes a machine, how should authorization be defined, verified, limited, and revoked?

This is also why many of the most critical actions in the industry today are not simply optimizing "payment capabilities," but are addressing the layer of authorization.

Visa's Trusted Agent Protocol essentially addresses the question of "who can act on behalf of whom." Visa officially defines it as a foundational framework that allows AI agents and merchants to communicate securely throughout the transaction process, with the aim of establishing trusted relationships in agent-driven commerce. The Asia-Pacific region's Visa Intelligent Commerce statement also emphasizes that the core value of the Trusted Agent Protocol is to enable merchants to identify and verify "trustworthy agents with genuine commercial intent."

Google's AP2 is also not merely a tool for "putting payments on agents." In its official description, Google defines AP2 as an open protocol that allows AI agents to securely initiate and execute payments on behalf of users, and it explicitly emphasizes that it must work in conjunction with agent protocols like A2A and MCP. In other words, it is not dealing with a single payment itself, but rather how agents can complete payment actions in a cross-system environment in an auditable and verifiable manner.

PayPal's recent consumer-facing explanation of agentic commerce has also straightforwardly addressed the authorization issue: users can choose to let AI complete purchases on their behalf or require "confirmation before purchase." This statement seems simple, but it has already clarified the layered authorization points of Agentic Payment—future payments will not all follow the same authorization model, but will simultaneously exist under both "incremental confirmation" and "pre-authorization" logics.

Stripe's MPP is similar. In its official blog, Stripe defines MPP as an open standard that allows agents and services to coordinate payments programmatically, supporting scenarios like micropayments and recurring payments. This definition sounds like it is discussing payment protocols, but what it is truly promoting is something else: transforming "payment" from a one-time human action into a system behavior that can be triggered by programs but must be constrained by rules.

So if we break down Agentic Payment further, we find that its real threshold is not the payment rail, but the authorization rail. In other words, what will truly hold value in the future may not be the "AI payment button," but the following capabilities:

First, authorization boundaries must be productized.

Users or enterprises cannot simply say, "Let AI buy for me." The system must articulate authorization as clear rules: amount limits, merchant whitelists, category restrictions, time windows, budget pools, whether automatic renewals are allowed, and whether secondary confirmation is required. These are not just experience details; they are the prerequisites for whether payments can truly be delegated.

Second, authorization status must be dynamically verifiable.

Today's payment authorizations are mostly one-time, but agent transactions are often continuous, cross-step, and even conditional branching. This means that in the future, the system must not only know "whether there is authorization," but also "whether this step is still within the authorization boundary."

Third, authorization must be revocable and traceable.

This is the biggest difference between Agentic Payment and ordinary automation. Because machines may continuously execute multiple actions while you are not watching, users and enterprises must be able to pause, retract, or modify permissions at any time. Otherwise, authorization will not be an efficiency tool but a source of risk.

Fourth, authorization results must be mutually understood by merchants and payment networks.

This is also why players like Visa, Google, and Stripe are all developing their own protocol layers. Because relying solely on users to set rules is meaningless; what is truly important is that merchants, wallets, payment networks, and clearing networks must be able to understand this authorization and decide whether to accept the transaction based on it.

From this perspective, Agentic Payment is not simply automating "human payments" but reprogramming the matter of "payment authorization."

This is also why I believe that what is most worth watching next is not which company has released a new payment protocol, but which company first turns the authorization system into a truly usable product.

Because protocols solve the question of "how to pay," wallets solve "who holds the account," stablecoins solve "what money to pay with," but what ultimately determines whether Agentic Payment can be established is: whether the system dares to allow machines to move money on your behalf without clicking confirmation. And this question, at its core, is not a payment issue. It is an authorization issue.

Six: The Endgame Is Clear, But the Real Deciding Factor Is Who Paves the Way First

The endgame of Agentic Payment is no longer scarce today. From the launch of the Tempo mainnet, the introduction of MPP, to Visa's Trusted Agent Protocol and Intelligent Commerce, and then to Google's AP2 / UCP, and Coinbase's Agentic Wallets, the industry has increasingly articulated what the future will look like: the initiators of payments will gradually shift from "manual confirmation by humans" to "machines executing within authorization boundaries"; and the layers of protocols, trust, wallets, and currencies surrounding this matter have also begun to emerge.

But the clearer the endgame, the more it indicates one thing: the real differentiator will not be who presents a more complete vision of the future, but who first successfully navigates the path in between.

Because payment is not an industry that can be rewritten solely by imagination. What it ultimately competes on is not demos, not launch events, not which terms are updated, but whether there are real transactions, whether there are continuous payments, whether they are accepted by users and merchants, and whether they are allowed by the system to execute stably without clicking confirmation. Coindesk's report on x402 in March has already highlighted this reality: technology is advancing, but real demand and large-scale adoption have not fully kept pace.

Therefore, what is most worth watching next is not who has released a new protocol, nor who has shouted a bigger vision for agentic commerce. What is truly worth watching are three more specific things:

• First, which type of real transactions will stabilize first.
• Second, which authorization system will be accepted first by users and merchants.
• Third, which layer of infrastructure will first transform from "conceptual puzzle" to "default component."

This is also why I believe that in the short term, the most important thing about Agentic Payment is not whether the endgame is correct, but whether the path is valid. The endgame is certainly important, but it is more like a direction.

What truly determines industry differentiation are always the details within the path:

• Who first turns authorization into a product,
• Who first turns trust into a standard,
• Who first turns wallets into accounts,
• Who first truly makes stablecoins usable by machines,
• Who first turns a certain type of transaction into a continuously occurring payment behavior.

Ultimately, Agentic Payment is not a single-point innovation.

It is more like a recombination: transforming money back into callable infrastructure, reconstituting payment into orchestratable system behavior, and redefining authorization into rules that machines can understand. And the true watershed of this matter lies not in "whether AI will pay," but in: when the financial system will be willing to acknowledge that machines can spend money on behalf of people within boundaries. Therefore, the endgame is clear to everyone. What truly decides the outcome is still who paves the way first.