Agent Wallet

An agent wallet is a cryptocurrency wallet controlled by software, usually an AI agent, that can hold funds and sign transactions on its own within limits set by a human or company owner. Instead of a person approving each payment, the agent decides when to pay, swap or move assets as part of a task, such as buying data, paying for computing power or settling a purchase for its user.

Agent wallets are a building block of agentic payments, in which software acts on behalf of people or businesses. They draw on existing crypto wallet technology, especially programmable smart accounts, and are often paired with stablecoins so that agents can pay in stable units of value. The design challenge is to give agents enough freedom to be useful while keeping the owner firmly in control.

Think of giving a trusted assistant a prepaid card with a spending limit. They can pay for taxis and lunches without phoning you each time, but they cannot spend more than the card allows, and you can cancel it at any moment. An agent wallet is that card for a computer program.

The program holds a wallet, follows rules you set, such as "spend no more than $50 a day" or "only pay these three services", and records every payment on a blockchain where you can check it. If it misbehaves or is hacked, the limits are meant to stop the damage from growing.

Several technical approaches are used, often in combination:

• Smart contract accounts. Wallets built as smart contracts, for example under the ERC-4337 standard described in account abstraction, can enforce spending caps, allow-lists and time limits in code.
• Session keys. The owner grants the agent a temporary key with narrow permissions that expire automatically.
• Key splitting. Multi-party computation wallets split the signing key between the agent, the owner and sometimes a provider, so no single party can move funds alone.
• Policy engines. Off-chain services check each proposed transaction against rules before co-signing.
• Mandates. Signed instructions, such as an agentic mandate, record what the user has authorised the agent to do.

Agent wallets are being tested for tasks such as paying per request for data and software services, buying cloud computing for other agents, managing on-chain positions, and handling micropayments too small for card networks to process economically. Developers have created toolkits that let AI models create and use wallets, and protocols that let a website request payment from an agent automatically. One example is x402, launched by Coinbase in 2025, which uses the long-unused HTTP 402 "Payment Required" status code to let services ask agents for stablecoin payment within a normal web request. Payment industry frameworks such as the Agent Payments Protocol aim to connect agent wallets with card and bank payments as well.

Agent wallets can be custodial, with a provider holding keys on the owner's behalf, or non-custodial, with keys controlled by the owner and agent. The choice affects who is responsible when something goes wrong and which rules apply. Providers that hold funds for customers may fall within money transmission or crypto licensing rules, and they still need to know who ultimately owns the wallet for know your customer and sanctions purposes. Regulators have not yet settled how existing payment rules, such as liability for unauthorised transactions, apply when a payment is made by software acting under a general instruction.

Giving software control of money creates new attack surfaces. Agents can be tricked by malicious prompts or websites into sending funds, a risk known as prompt injection. Bugs in the agent's logic or in the wallet contract can drain balances quickly, and an agent can make costly mistakes at machine speed. Keys stored on servers are exposed in the same way as any hot wallet. Good practice includes tight spending limits, allow-lists of approved counterparties, monitoring and alerts, and the ability for the owner to freeze the wallet at once.