Tencent's Agent wallet that is currently in the testing phase has been preempted by Amazon.

Do you still remember Tencent's exclusive AI card?

On June 17, 2026, WeChat Pay publicly introduced the "AI Exclusive Card" that was under testing. In short, this is a restricted payment card specially designed for AI Agent use. Users can top up in advance, set a limit, and then bind it to agents such as WorkBuddy. After that, the AI can complete searches, recommendations and order placements according to user instructions, and the fees will be deducted from this exclusive card.

Martin Lau, President of Tencent, mentioned in the earnings call in November 2025 that "WeChat will eventually launch AI agents that can help users complete a large number of tasks within WeChat. WeChat has a social ecosystem, a mini-program ecosystem, and a seamlessly connected payment ecosystem. The ideal assistant understands user needs and executes all tasks within the WeChat ecosystem."

However, up to now, this payment sandbox has not been launched, and no follow-up news has been released since then.

But news came out today that Amazon has launched this product first.

On August 19, Amazon AWS announced that Amazon Bedrock AgentCore Payments has officially entered the GA stage. Previously, there were only preview functions for some simple payment scenarios. Now it allows AI Agents to independently purchase paid APIs, MCP services, digital content and model inference capabilities under user authorization and budget constraints.

Users can top up their wallets, and developers' keys are placed in AgentCore Identity Secrets Manager. Agents do not need and cannot directly see the original credentials. The upper limit of the amount and expiration time can also be set for each payment. The infrastructure will perform deterministic verification before the transaction is signed to prevent the Agent from overspending beyond its authority.

This coincided with the function of WeChat Pay's AI Exclusive Card that Tencent had previously tested.

Tencent's approach is also to not give Agents full wallet permissions for users, but to allocate a separate sum of funds. Users top up first, set the limit, then WorkBuddy completes searches, recommendations and order placements according to instructions, and finally the user confirms the payment.

However, there are still differences between the two. Although on the surface, both Tencent and Amazon are letting AI spend money on behalf of users, looking deeper, you will find that Tencent is more focused on life scenarios, while Amazon is more business-oriented.

To see the difference between the two clearly, we have to start with their respective service targets and payment methods.

What is the difference between Amazon Agent Wallet and Tencent's AI Exclusive Card?

Let's look at Amazon first.

AgentCore Payments is not aimed at ordinary consumers, but at enterprises and developers deploying AI Agents. What it solves is to allow Agents to directly purchase the digital resources they need when performing tasks, such as paid APIs, MCP Servers, real-time data, web content, model inference and computing power services.

For example, a research Agent can purchase financial data on a per-use basis, a programming Agent can call paid tools, and a browser Agent can access paid web pages.

Therefore, rather than being automatic payment, I think AgentCore Payments is more like a set of payment infrastructure deployed for developers.

Developers first choose wallet service providers such as Coinbase CDP, Stripe or Privy, prepare the corresponding API Key and Secret, then hand over the credentials to Amazon's AgentCore for management, create a payment manager, payment connector and payment tool, and then create a payment session for each Agent interaction.

If it is the end user of Amazon AgentCore Payments, they need to top up their wallet and explicitly authorize the Agent to spend money on their behalf, which is similar to the AI Exclusive Card.

After that, when the Agent encounters a paid service that supports x402 or MPP, it can initiate a payment request. AgentCore will check the amount, session status and budget upper limit before the transaction is signed. Only when the rules are met can the underlying wallet be called to complete the signature and payment.

Overall, the key point is not that "a card is arranged for AI", but that payment is embedded into the Agent's workflow.

To sum up, if my workflow needs to use a certain service that requires payment, AgentCore Payments will complete the payment action in the background for me, so that I do not need to jump out of the workflow to make the payment.

Then let's look at Tencent. As mentioned at the beginning, users need to activate the AI Exclusive Card in WeChat Wallet first, transfer a sum of budget from WeChat Change to it, and then bind it to Agents such as WorkBuddy.

Then when the user puts forward demands such as "help me buy a group coupon" or "help me book a service" in the conversation with the Agent, the Agent is responsible for understanding the demand, searching for products, comparing options and generating orders.

Next, WeChat Pay is responsible for sending the order and amount to the user's mobile phone. After the user confirms and verifies the password, the funds will be deducted from the AI Exclusive Card.

Therefore, the AI Exclusive Card of WeChat Pay is essentially a restricted secondary card specially designed for Agent use. Moreover, unlike AgentCore Payments, the AI Exclusive Card cannot connect to other wallets, and can only consume the funds in the AI Exclusive Card.

In other words, Tencent restricts the AI's spending power to the question of "how much money is in this card", and first turns the most dangerous unlimited authorization into a visible, controllable and revocable balance limit.

However, its AI Exclusive Card is more life-oriented than AgentCore Payments.

Tencent has previously disclosed several use cases, one of which is that the user opens WorkBuddy, summons the "Meituan Life Assistant" on the expert page, and asks it "what good group-buying deals are nearby".

WorkBuddy will combine the user's location information and preferences to search for merchants, and then recommend products.

The key point of the AI Exclusive Card link is to compress the actions that were originally scattered in several applications such as search, price comparison, order placement, and opening WeChat Pay into one conversation. The AI runs errands for the user first, and WeChat Pay collects the payment later. Payment is still the last step of the entire consumption closed loop.

The biggest difference between the two is that the two companies want to place payment in different positions. Tencent wants to advance AI from "being able to recommend" to "being able to consume", so that after the user says a word, life services can form a closed loop within the WeChat and mini-program ecosystem.

Amazon hopes that payment will become a basic action in the Agent's workflow. It does not require developers to re-embed the checkout page, subscription system and API Key of each service provider into the Agent. Instead, it allows the Agent to actively discover paid services in the background, complete settlement, and then continue to perform tasks.

This is why it seems that both companies are making AI wallets, but in fact they have moved towards two different markets.

Compared with WeChat Pay, what are the key designs of AWS?

The reason why Amazon dares to let Agents complete the consumption action on behalf of users actually lies in three layers of key designs.

The first layer is the separation of identity and credentials.

The AI Exclusive Card of WeChat Pay does not need to give the user's payment password to WorkBuddy. The Agent can only complete the payment through the binding relationship and user confirmation. The user needs to click confirm, and its payment action is carried out "one by one".

In contrast, Amazon's approach is more suitable for enterprise environments. Developers' Coinbase or Privy keys are hosted by AgentCore Identity Secrets Manager, and the Agent itself cannot get the original keys.

It can only call the payment tool and submit a payment request to AgentCore. After that, AgentCore uses a short-term token derived from the credentials to instruct the wallet service provider to perform operations such as creating a wallet or signing a transaction.

In other words, the Agent only has the ability to "request payment", but does not have the ability to directly transfer money with the private key.

In other words, AgentCore Payments never touches the money from beginning to end, it is only responsible for passing messages between the two sides, such as "the money has been transferred over there" and "the money has been transferred here", nothing more.

The AI Exclusive Card works like "wait a minute, I will take the money out of my pocket and give it to you".

In short, Tencent focuses on solving the problem of "how much money can I spend", while Amazon focuses on solving the problem of "in what identity, with what permissions, and under what conditions this Agent initiates this payment".

For a single consumer scenario, balance isolation is intuitive enough. But for enterprises, the same Agent may serve many users, call many suppliers, and need to distinguish budgets between different sessions.

If it is only given one wallet address, problems such as mixed accounts, unreachable permissions, and difficult audits will soon be encountered later.

The second layer is payment sessions.

Amazon wrote in its documentation that AgentCore Payments treats payment as context.

In addition to insufficient funds, if a single session takes too long, or the session content is problematic, the Agent cannot continue to make payments.

For example, among the many scenarios where AgentCore Payments can be used, there is a research scenario. The user wants to study physics knowledge, and the Agent is responsible for purchasing relevant literature materials for the user. But if a payment request unrelated to physics appears, the Agent cannot complete the payment either.

The third layer is deterministic verification.

Large models are non-deterministic systems. They may misread a sentence on a web page as payment authorization, or repeatedly retry after the interface times out.

It is often said that "a hero's origins need not be inquired about", but Amazon explicitly places the budget verification at the infrastructure layer before the transaction is signed. Amazon believes that no matter who you are, your origins must be clearly checked.

Only after the verification is passed can the underlying wallet sign. This order is very important, because it means that Amazon does not hand over the last safety gate to the model. The model only has one power, that is, to tell the wallet, "I really need this sum of money".

The fourth layer is protocol orchestration.

At present, WeChat Pay's AI Exclusive Card is mainly embedded in the order process of WeChat Pay and life services, but Amazon needs to face a large number of potentially incompatible machine payment protocols.

During the preview phase, AgentCore Payments supported X402 initiated by Coinbase. Amazon said that after entering GA, MPP jointly developed by Stripe and Tempo has also been added.

Amazon's approach is actually a bit like Claude Code. No matter who the underlying service provider is, there is no need to rewrite the Agent for each payment standard. After accessing it, it will run directly according to the set of rules that Amazon has already orchestrated.

The fifth layer is the "up to maximum" scheme. Traditional pay-per-use usually requires confirming the price before the request, but for model inference, computing power and data query, the final consumption is often unknown in advance.

A single request may only generate hundreds of tokens, or consume more resources as the context becomes longer. Amazon added an "upto" mechanism to x402, allowing the Agent to authorize a maximum amount first, and the service provider settles according to the actual consumption after the call ends. In this way, service providers can charge according to the actual number of tokens, computing power or the amount of returned data, instead of setting a fixed price for each request.

The AI Exclusive Card uses fixed payment, which means you pay exactly the price marked for the product, such as the marked price of a cup of milk tea.

In Amazon's official demonstration, Agents can dynamically switch between different models. Simple tasks call cheaper models, complex tasks call more powerful models, or switch to another inference service when the price changes.

The Agent may even compare the prices and effects of different models during the execution of the task, and then decide who to assign the next call to.

After such a set of operations, it can help users save a lot of money.

Why doesn't Tencent dare to make AI wallets into infrastructure?

The AI Exclusive Card also uses the protocol called X402, but its full name is Tencent X402 Agent Pay, which is a completely self-developed, domestically closed-source interface within the WeChat Pay system, and is not the same thing as the international X402 mentioned earlier.

Just because the underlying logic is similar, both are based on the Http 402 status code to trigger instant micro-payment when the Agent calls the API, so similar names are used.

As early as April 2026, Tencent officially released the WeChat Pay access Skill, which is positioned as a developer tool for AI-IDEs that support the Skill protocol such as Cursor, to help developers select payment products, perform access security scans, and troubleshoot errors. It is not a function that ordinary users can directly perceive.

In June, it iterated to version 2.0, adding code security assessment and dynamic troubleshooting, and updating the official documentation at the same time. In the June update, the AI Exclusive Card was also released. It was first launched for the WorkBuddy Mac version, used to run group-buying order scenarios for Agent Skills such as Meituan Life Assistant.

However, the GitHub page of the WeChat Pay access Skill shows that the repository has not been updated since June.

Tencent can make AI payments, but for the time being, it does not dare to make it a cross-platform, automated open wallet infrastructure for the machine economy like AgentCore Payments.

As mentioned earlier, the AI Exclusive Card still requires the user to click to confirm at the last step. Then imagine, what if Tencent changes this process to "automatically pay as long as the limit is met"?

The problem will immediately become complicated. If the AI misinterprets a sentence from the user, does it count as user authorization? If malicious prompt words in the web page induce the Agent to buy something that should not be bought, should the platform compensate? If the Agent repeatedly calls due to interface timeout, resulting in 10 deductions for one order, should the responsibility fall on the Agent, the merchant or the payment institution?

The reason why Amazon dares to do this is that it limits automatic payment to scenarios of developers, enterprises and small-value digital services. The user base is small and has strong risk-bearing capacity. Except for hotel reservations, transactions are often API, content or inference calls ranging from a few cents to a few dollars. Even if an error occurs, it will not cause much loss.

Once the AI Exclusive Card is opened to hundreds of millions of ordinary users, the responsibility Tencent faces will be far beyond imagination. When the user scale is too large, the platform cannot rely on only one budget rule to solve all risks.

Tencent's advantage is precisely that WeChat Pay connects WeChat, official accounts, mini-programs, merchants and local life services, but that's all. All payments of the AI Exclusive Card can only be used around this ecosystem at present.

What Amazon is doing is cross-ecosystem payment. An Agent may call an API that does not belong to AWS when performing a task, then access paid content behind Cloudflare, then purchase a model inference through BlockRun, and finally call Travala's travel MCP Server to book a hotel.

For Tencent, if it wants to become the underlying wallet for such transactions, it has to handle cross-platform identities, cross-border payments, after-sales of digital content, and technical standards of different service providers. Therefore, Tencent's current approach is to call on merchants to settle in the WeChat ecosystem, rather than aligning all parties' protocols.

Most importantly, we have to admit that there are performance gaps between models. Although Amazon's own model Nova is weaker than Tencent's flagship model Hy3, Amazon's AgentCore Payments itself is not bound to a certain model. AWS has explicitly stated that it can work with models inside and outside Amazon Bedrock, including Anthropic's Claude, OpenAI's GPT, Google's Gemini and so on.

On Tencent's side, the publicly launched entry of the AI Exclusive Card is Tencent's own WorkBuddy, and WorkBuddy mainly uses domestic models such as Hunyuan, DeepSeek, GLM, MiniMax and Kimi.

Tencent is not unable to access third-party models. T