Why AI Agents Use Crypto and Not Banks
AI tools have so far mostly written text, answered questions, and searched the internet. The next step is for AI agents to perform the tasks themselves, including paying for the services they need. Instead of a chatbot waiting for a query, the agent sets a goal on its own, collects information, and performs actions, including financial transactions. During 2026, several major crypto and payment companies launched infrastructure that allows AI agents to hold and spend cryptocurrency without human approval for every transaction. Coinbase, Stripe, MoonPay, and Google are just some of the players that released their own solutions for so-called agent payments in the first half of the year. The question arises as to why crypto has become a natural choice for software that needs to be paid, instead of traditional banking.
Source: cointelegraph
Why can't AI agents go to the bank?
Opening a bank account requires KYC, or verification of the client’s identity: ID card, proof of address, sometimes tax number. An AI agent has none of that. He does not have a passport, cannot sign a contract and does not have legal personality. The banking system is built for people and registered companies, and an agent is neither. A crypto wallet works differently. It is not an invoice issued by an institution, but a cryptographic pair of keys. Generating a wallet requires nothing other than running a cryptographic function that produces a private key and a public address. There is no form, no identity verification, no institution that grants access. In addition to identity, there are two additional reasons. Card processing fees are 2 to 3 percent plus about thirty cents per transaction, which makes a one-cent payment meaningless because the fee exceeds the transaction itself. ACH transfers in the US take one to three business days, and bank transfers only work during business hours. An AI agent who pays for access to data or computing power needs settlement in seconds, not days.
Source: cointelegraph
What is an agent wallet?
An agent wallet is a crypto wallet made specifically for an AI agent rather than a person. An agent can use it to receive and send cryptocurrency, pay for the services of other agents or APIs, and execute transactions without a human approving each individual payment. The difference to a regular wallet is who is in control. With an agent wallet, the funds are under the control of the agent and the rules set by the person who started it, not necessarily the person who funded the wallet. If someone finances the agent’s wallet in order for the agent to pay for services on his behalf, these funds are under the agent’s control from that moment on, within the limits set by him.
Source: cointelegraph
Who builds the infrastructure for AI agents?
Several companies in 2026 launched concrete solutions:
● On February 11, 2026, Coinbase launched Agentic Wallets, wallets protected by MPC technology, i.e. cryptography shared between multiple parties, with programmable spending limits per transaction and per period, with no gas fee on Base. The wallet is installed via the command line or as an MCP server compatible with Claude, Codex, and Gemini tools, so AI models can use it directly
● The x402 protocol, launched by Coinbase in May 2025, reuses the HTTP status code 402 Payment Required from the original HTTP specification, which had no real application for decades, to embed the payment stablecoin directly into web requests. Cloudflare founded the x402 Foundation with Coinbase, and the protocol has processed more than 50 million transactions so far●
In September 2025, Google announced AP2, an open protocol for agent payments that supports cards and bank transfers in addition to crypto, along with Coinbase, Ethereum Foundation, and MetaMask. More than 60 partners have tied themselves to AP2, including Mastercard, PayPal, American Express, and Salesforce
● Stripe launched Machine Payments on Base on February 11, 2026, based on the same x402 protocol and integrated into the existing Stripe API
● MoonPay launched MoonPay Agents on February 24, 2026, a non-custodial layer in which the user passes KYC verification once, and then the agent independently trades and transfers assets on Ethereum, Solana, Base, Arbitrum, Optimism, Polygon and Bitcoin
The common denominator of all these solutions is the stablecoin as a means of settlement and the blockchain as the layer on which transactions are confirmed. All of this is already in practice: CoinGecko charges access to market data one cent in USDC per query without any account, and Google and OpenAI have introduced agent payment directly at checkout when making a purchase.
Source: cointelegraph
What are the risks?
When an agent spends money without human approval for every transaction, three specific problems open up. An agent could unknowingly pay for an address that is on the sanctions list. It could exceed the spending limits set for it. Or someone could use malicious instructions to manipulate the agent into making a transaction that the owner didn’t want. Traditional KYC and AML anti-money laundering systems assume that there is a responsible person for each transaction. With autonomous software, this model does not coincide, so new tools are being developed: spending limits built directly into smart contracts, lists of approved addresses with which an agent is allowed to do business, KYT checks, i.e. checking the transaction itself instead of just the client, and complete records of each action for later audit. The funds that someone transfers to an agent wallet are under the control of that agent and the platform that runs it. This is the technical difference of who actually manages the funds once they are transferred to the agent wallet.
Source: cointelegraph
Where does this lead?
AI agents today can independently pay for an API call, GPU hour or data query, usually in stablecoin and with preset spending limits. According to McKinsey’s estimates, such an agent trade could reach $3 trillion to $5 trillion globally by 2030, but this is an estimate by a research firm, not a guaranteed outcome. What does not yet exist is a uniform standard for how much autonomy an agent is allowed to have and who bears responsibility when something goes wrong. For now, a simple rule applies: the agent manages the funds intended for him, within the limits set for him. Control over the remaining assets still remains with the person holding the private keys.