On August 4, Cloudflare made a highly symbolic move by unveiling Cloudflare Wallets—infrastructure designed to give AI agents wallets of their own.
With this infrastructure, AI agents are no longer limited to calling APIs, reading data, or executing code. They can also have independent virtual wallets and use stablecoins such as USDC to purchase APIs, data, content, and computing services within predefined budgets and permission boundaries.
A different approach was gaining momentum in the same week.
On August 6, MetaMask also rolled out agent wallet, allowing agents to connect to on-chain wallets and execute swaps, trade perpetuals, participate in prediction markets, and manage liquidity within permissions configured in advance by users.
The two products may appear to serve different purposes, but together they provide a critical piece of infrastructure that AI agents have long lacked. This could be one of the most important structural developments to emerge from the recent convergence of AI and crypto—and one worth watching over the long term.
1. Agents Have Long Lacked a Way to Pay on Their Own
Consider Cloudflare first.
Although the capabilities of individual agents and multi-agent systems have advanced significantly this year, they can still become stuck in a very traditional process: find a service, visit its website, create an account, add a credit card, purchase a subscription, obtain an API key, and only then begin using the service.
For a human, this process is merely inconvenient. For software trying to complete a task autonomously, however, any step involving login, registration, payment, or identity verification may force it to stop and ask a human to take over.
In other words, the “brains” of agents have advanced rapidly over the past several years, but the internet’s payment infrastructure is still fundamentally designed for humans.
This is precisely what x402 seeks to change.
It revives the long-standing but rarely used HTTP status code 402 Payment Required, embedding payment requests directly into the internet’s basic request-response flow.
Under Coinbase’s x402 design, when an agent requests access to a paid API, the server can tell it directly how much to pay, which assets it accepts, and where the payment should be sent. The agent completes the payment and retries the request with proof of payment. The server verifies the payment and returns the requested resource.
The original process—
“Create an account → connect a payment method → add funds or purchase a subscription → obtain an API key → call the service”
—can therefore be compressed into:
“Send a request → receive payment instructions → pay → access the resource.”
This flow may not require an account or subscription. An agent would also no longer need to purchase a monthly or annual plan simply to make a handful of API calls.
Removing a few steps may not sound transformative, but it is particularly well suited to AI agents.
The payment mechanism an agent truly needs is not one that requires its owner to stop and enter a verification code before every purchase. It needs a payment protocol that software can understand and execute automatically, just like any other programmatic call, while supporting precise usage-based pricing.
Stablecoins provide an ideal settlement layer for this model.
Cloudflare’s current x402 developer documentation already supports machine-to-machine payments using USDC and other on-chain assets. An agent can pay directly when requesting an API, MCP tool, or other digital resource instead of being redirected to a traditional checkout page. Further reading: “Crypto AI Protocol Landscape: Building a New Operating System for AI Agents on Ethereum”
This could even change how internet content is priced.
The traditional internet generally offers two choices: provide content for free or place it behind a subscription wall that requires a human to register and become a paying member.
For agents, a different model may be more natural. Instead of requiring them to subscribe to an entire service, providers could charge for the data consumed, API requests made, compute used, or content pages accessed.
This is what makes Cloudflare Wallets worth watching.
A research agent, for example, could receive a budget of 10 USDC and independently compare the price, speed, and quality of dozens of data sources. If an API call costs only a few cents, the agent could try the service immediately. If the result is unsatisfactory, it could move on to the next provider without asking its owner to approve every few cents of spending.
Interestingly, these restrictions may appear to constrain the agent, but they are precisely what gives it greater autonomy.
If a user must manually approve every 0.01 USDC request, the supposedly autonomous agent remains little more than a partially automated tool.
Only when the user first establishes a sufficiently clear budget boundary—retaining ultimate control outside that boundary—can the agent operate freely within it.
The launch of Cloudflare Wallets therefore reflects a deeper shift.
Historically, internet infrastructure assumed that the participants in economic transactions would mainly be individuals and businesses.
Now, from identity and payments to pricing models, part of that infrastructure is being deliberately redesigned for another category of participant: AI, or more specifically, the agent itself.
2. How Can Agents Be Given Economic Autonomy?
If Cloudflare is primarily addressing how agents can purchase services, MetaMask Agent Wallet takes the next step by asking how agents can use assets directly.
It seeks to let agents perform on-chain operations within permissions predefined by users. This is fundamentally different from asking an AI to analyze whether ETH is worth buying.
Historically, the division of labor between humans and AI was relatively straightforward: the AI gathered information, analyzed the problem, and made a recommendation. The human then decided whether to act.
With an agent wallet, an instruction could gradually evolve into something like:
“If ETH falls to around $3,000 while gas is below its 24-hour average, buy 0.2 ETH.”
The user provides the objective, conditions, and permissions. Continuous monitoring, condition evaluation, transaction preparation, and even final execution can then be delegated—in whole or in part—to the agent.
This is the layer that makes an agent’s “economic autonomy” genuinely significant.
It does not mean that the agent owns property in its own right. Rather, it gains an account, a discretionary budget, and a set of economic permissions it can invoke as circumstances change.
It can independently purchase external information and computing resources. It can also deploy real assets to accomplish objectives within rules defined by the user.
This step may seem like a natural progression, but it also connects AI errors directly to real economic losses for the first time.
If a conversational AI misunderstands a sentence, the usual result is simply an incorrect answer. Once it has a wallet and execution authority, the same misunderstanding, prompt injection, or malicious tool call could immediately result in an irreversible on-chain transaction. Further reading: “A Signature Is More Than a Signature: When an AI Agent Signs for You, Who’s in Control?”
This is why an unrestricted wallet cannot simply be handed to an AI.
Users can define daily spending limits, permitted protocols, and risk preferences in advance. Supported EVM transactions can also pass through transaction simulation, threat scanning, and MEV Protection.
If a transaction is identified as abnormal or falls outside the user’s predefined policy, the system pauses automated execution and requires the user to reauthenticate via 2FA before proceeding.
An important principle is therefore becoming clear: economic autonomy does not mean unlimited authorization.
A genuinely useful agent is more like an employee with a corporate card and clearly defined responsibilities than someone handed the keys to the company vault.
What it can buy, how much it can spend in a single transaction or over one day, which decisions it can make independently, and which actions require renewed approval should all be established before authority is delegated.
From this perspective, the central innovation of an agent wallet is not merely giving AI a wallet. It is beginning to address systematically how humans can delegate economic authority to software without surrendering control.
Once we reach this point, the wallet itself must also change.
3. What Kind of “New Wallet” Is Needed When Agents Can Spend Independently?
For more than a decade, the central challenge for crypto wallets has remained remarkably stable: how to manage private keys securely.
No matter how wallet interfaces have evolved, the underlying relationship has remained the same. A human initiates the operation, reviews the transaction, and provides the final signature. The wallet’s most important responsibility is to protect the private key that determines asset ownership and grants final authorization.
The arrival of agents introduces another layer into this process.
Users may no longer need to construct every transaction themselves. The interaction model is gradually shifting from “a human operates the assets directly” to “a human defines an objective and delegates part of the execution authority to an agent.”
This means future wallets must answer an entirely new set of permission questions:
- Who can use assets on my behalf?
- Which assets and protocols can it access?
- What is the maximum value of a single transaction, and how many transactions can it execute per day?
- Which decisions can it make independently, and which require renewed confirmation?
- If the agent behaves abnormally, can I immediately pause it, revoke its permissions, and recover the assets?
For ordinary users, the key question is how they can confidently manage agents with real execution capabilities.
This is another question imToken is exploring through UI 3.0 and its thinking around agent Wallets.
In imToken’s vision for next-generation wallet interactions, one important change is that the user’s role begins to shift from “Operator” to “Manager.”
Today, completing a transfer requires the user to select a network, enter an address and amount, assess the gas fee, review each step, and sign the transaction.
In an intent-driven wallet, the user may only need to say:
“Send Frank 500 USDT.”
The system could translate that natural-language instruction into a structured intent, identify the recipient, amount, asset, network, and estimated fee, and then present the structured transaction details to the user for final confirmation.
Reducing the number of visible steps does not mean that the boundaries of control can disappear with them.
On the contrary, as more of the execution process becomes automated, the steps no longer visible to users must be governed by clearer authorization mechanisms.
Under imToken’s Agent Wallet design concept, each agent granted execution authority corresponds to a separate agent account. Its session key is generated and isolated within a trusted execution environment, or TEE, and never leaves that secure environment.
At the same time, every agent account must be bound to an explicit policy covering protocol allowlists, per-transaction limits, daily limits, operating frequency, validity periods, and other restrictions.
The agent therefore does not receive an unrestricted wallet. It receives an execution account enclosed by policy guardrails.
Within this relationship, the user retains a higher level of control. The user can adjust the policy, pause or resume the agent, revoke its permissions, and recover the funds at any time.
AI can help interpret intent, plan execution paths, estimate fees, and identify risks. But the agent can execute only within the policy authorized in advance by the user. Any operation outside those boundaries must return to the user for confirmation.
More importantly, authorization should not be a single leap from zero to complete control. It should expand gradually as trust develops.
A newly adopted agent might begin in an observation and analysis role. As the user develops confidence in it, the agent could be allowed to make recommendations and prepare transactions. The next stage could permit execution after user confirmation. Only when the rules are sufficiently clear and the risks controllable should the agent be allowed to execute automatically within a predefined strategy.
The progression from L0 observation to L1 recommendations, L2 execution after confirmation, and L3 autonomous execution within policy represents a gradual trust model.
An agent’s autonomy is not something the system should grant by default. It is authority that the user confers one step at a time.
This could also change why users open their wallets.
Today, people generally open a wallet because they want to check a balance, make a transfer, or execute a swap.
Once agents take over more routine operations, users may open their wallets with a different question in mind:
“What has my agent done recently, and is there anything that needs my attention?”
The wallet’s primary interface may therefore evolve from a transaction interface into a management dashboard.
At that point, a wallet would no longer serve only as a place to store assets and initiate transactions. It would function as a permissions and control layer between users and their agents.
This may represent an important shift in the wallet’s value proposition in the agent era—from “securely manage your private keys” to “securely manage your assets and the agents you authorize to use them.”
Final Thoughts
Viewed over a longer time horizon, giving AI agents economic autonomy may be one of the most important structural developments to emerge from the convergence of AI and crypto.
Smarter models and agents capable of using more tools still operate primarily within the information world. Once an agent has an account, a budget, and the authority to move and deploy real assets, however, it becomes an active participant in real economic activity for the first time.
The factor that ultimately determines whether agents can enter real financial environments at scale may therefore no longer be simply whether they are intelligent enough.
The more important question is whether we can build a permission system that matches their capabilities.
The other side of autonomy is always authorization.
Wallets may not simply disappear into the background in the agent era. Instead, they may take on an even more important role: enabling genuine automation while ensuring that ultimate control always remains with the user.
After all, allowing an agent to act freely does not mean that humans must surrender control.
That boundary may be the central question wallets need to answer before economic autonomy can become a practical reality.