Circle Internet Financial
Circle Internet Financial Logo

Nov 19, 2026

November 19, 2026

How AI Agents Hold and Spend Money

what you’ll learn

Discover how AI agents hold and spend money using stablecoins, AI agent wallets, spending limits, delegated permissions, and session keys.

Learn how AI agent wallets work, why stablecoins fit agentic payments, and how spending policies, delegation, and session keys give developers control over automated purchases.

How AI Agents Hold and Spend Money

Imagine giving a coding agent a $20 test budget that allows it to buy the resources it needs while work is underway. Its operator, the person or team that deploys and oversees the agent, needs confidence that it will stay within budget, use approved providers, and return a record of each purchase.

Making that possible is an infrastructure problem. AI agents can hold and spend money through wallets or payment accounts operated under delegated authority, meaning the operator grants the agent limited power to act on the operator's behalf. For developers building AI agent payments, the essential components are a funded balance, protected signing credentials, enforceable spending rules, and reliable settlement. These components are what turn the agentic economy’s promise of delegated, automated work into safe and more controlled financial workflows for operators and builders.

Below, we’ll revisit the coding agent example as we explain how AI agent payment infrastructure works, what its components are, and how human operators can manage agent-led payments.

What is an AI agent wallet?

An AI agent wallet is infrastructure that lets an agent access funds, authorize transactions, and receive payments on behalf of an operator. The funds themselves are recorded on a ledger, while the wallet supplies the tools and permissions to use them. In an onchain AI payments system, meaning one that records transactions on a blockchain, that ledger is the blockchain itself, and each balance is tied to the AI agent wallet's address, a public identifier that works much like an account number. Giving an agent access to a wallet doesn’t make the agent an independent legal owner of the funds; it still acts on the operator’s behalf.

For agents, the practical challenge is using a wallet without a person signing off on every transaction. Smart contract wallets and embedded wallets are two complementary ways to solve that challenge.

What are smart contract wallets?

Smart contract wallets hold assets at an address governed by code, specifically a smart contract, which is a program that runs on a blockchain. That code’s authorization logic can require multiple approvals, restrict certain transactions, or support account recovery. Account abstraction, which lets a wallet run on custom rules written in code, makes these programmable wallets practical to use. It also enables features such as sponsored transactions, where an app or other party covers network fees, and batched actions, where multiple operations are bundled into one transaction.

Wallets also depend on a private key: a secret credential that approves, or signs, each transaction. Anyone who holds the key can move the funds, so how that key is stored and used, known as key management, is central to wallet security. For AI agents, this creates a problem. Handing the key directly to an AI model would give it unrestricted access to the funds and put the key itself at risk of exposure.

What are embedded wallets?

Embedded wallets address this challenge by building key management into the application. The key stays inside a protected signing service, and the agent asks that service to approve transactions without ever seeing the key. For example, Privy’s wallet security architecture splits keys into encrypted pieces that are stored separately and performs signing inside secure enclaves, which are isolated computing environments that shield sensitive operations.

The two wallet types work together: embedded signing protects the key, while smart contract rules govern what that key is allowed to do. With both in place, developers can set firm boundaries on what an AI agent wallet can do.

Why bank accounts alone fall short for AI agent payments

If wallets take this much purpose-built infrastructure, why not give an agent access to a bank account instead? Conventional bank onboarding is designed to identify people: account holders and those authorized to act for them, such as a business’s owners and representatives. Payment providers run similar identity checks. For example, Square’s identity verification requirements ask US sole proprietors to provide personal details such as their legal name, date of birth, and Social Security number or Individual Taxpayer Identification Number. Agent software can’t go through that process itself, so it typically acts through its operator’s account and permissions. That leaves developers to translate the operator’s broad financial authority into narrowly scoped access for each agent.

Payment timing adds more friction. ACH payments, which move money between US bank accounts over the Automated Clearing House network, can settle on the same business day, the following day, or up to two business days later. Because settlement runs on that schedule instead of completing when an API request does, ACH is ill-suited for AI agent payments.

Banking services may offer account APIs and 24/7 instant payments through services like FedNow, the Federal Reserve's instant payment service, but conventional accounts alone still do not support agent-specific permission models. Developers still need the building blocks to connect access, spending policies, and payment confirmation to the agent’s workflow. Programmable wallets with enforceable onchain spending policies provide those building blocks, and stablecoins supply the money those wallets move.

Why stablecoins fit agentic payments

For agents buying services that accept onchain payments, stablecoins paired with smart contract wallets offer a practical starting architecture. Stablecoins are digital assets designed to track a reference value, such as US dollars, euros, or other fiat currencies. In this setup, the wallet supplies programmable authorization, and a suitable blockchain supplies settlement, including outside banking hours.

This combination suits agentic payments because a single software workflow can both check the agent’s permissions and submit a transfer. Smart contracts can also hold funds in escrow, releasing them only when defined conditions are met, or divide payouts among multiple parties, as explained in Circle’s programmable money guide.

Choosing which stablecoin and network to use involves several trade-offs:

  • Provider acceptance: whether the service provider accepts that stablecoin on that network
  • Issuer risk: potential losses tied to the company that issues the stablecoin
  • Liquidity: how easily the stablecoin can be exchanged
  • Network fees: the cost of having the blockchain process and record a transaction
  • Finality: the point at which a payment can no longer be reversed

Fees matter most for lower-value payments, such as micropayments of a few cents or nanopayments of fractions of a cent, where total processing costs have to stay small relative to the payment itself. And if a provider ultimately needs funds in a bank account, converting stablecoins into bank deposits adds an extra step to the payment flow.

How to set spending limits for AI agents

Stablecoins and programmable wallets give an agent the means to pay. The next question is how to make sure it pays only for what it should. Financial controls need to operate outside the AI model’s reasoning. A prompt that says “stay within budget” expresses intent, but nothing enforces it. A wallet policy determines whether a transaction can proceed at all.

A useful spending policy should specify:

  • Budget: A total allowance and a maximum per transaction, including fees
  • Destinations: Approved recipients, networks, assets, and contract actions
  • Duration: When authority starts, expires, or must be renewed
  • Escalation: Which purchases require human approval

For our coding agent example, that could mean a $20 total allowance, payments only to approved providers, authority that expires when the job ends, and human approval for any purchase above a set threshold.

Controls for AI payments may run in smart contract code, a protected signing service, or both. Privy’s wallet policy documentation illustrates how infrastructure can restrict transfers and contract interactions.

Whichever approach they choose, developers should verify which transaction paths each rule actually covers. Token approvals, which authorize a smart contract to spend tokens from the wallet later, can grant future spending authority, so they need limits too. And when multiple agents draw from the same funds, they must share an aggregate budget so separate payment requests can’t each spend the same remaining allowance.

How session keys and delegated permissions control AI agent spending

A spending policy defines what an agent is allowed to do. Delegation determines how that authority passes from operator to agent in the first place. One common pattern gives each agent its own separately funded wallet. Another grants an agent limited access to an existing wallet. In either case, sensitive actions can require additional approval while routine purchases proceed automatically.

Session keys are a common way to put this delegation into practice. A session key is a signing credential granted specific permissions, usually for a limited period. The owner authorizes the credential, the wallet checks each request against its permitted actions, and then the agent can submit eligible transactions without receiving the owner’s primary key or requesting repeated signatures. When evaluating a session key implementation, such as Alchemy's session key APIs, look for clear permissions covering token allowances, contract functions, gas limits (caps on the network fees a transaction can consume), and expiration.

Revisiting our example once again, suppose the coding agent receives a session key that lets it spend up to $20 with an approved provider during a two-hour job. The operator controls authorization, renewal, and revocation. Together, session keys and delegation serve as important guardrails that constrains agent activity. An agent’s permissions should be configured so it cannot alter or expand its own permissions. 

How do AI agents pay for services?

With a funded wallet, a spending policy, and a session key in place, our coding agent is ready to make a purchase. Say it encounters a testing service priced at $2 per test run. Using its session key, the associated payment flow could follow these steps:

  1. Read the offer: Obtain the service terms, price, recipient, accepted asset, and network.
  2. Check authority: Validate the purchase against session permissions and the remaining budget.
  3. Authorize payment: Have the protected signer authorize the permitted payment.
  4. Confirm the outcome: Obtain payment status, receive the session, and check that access works.
  5. Record the expense: Connect the receipt and settlement reference to the task, provider, and authorization used.

Each step also needs a plan for when something goes wrong. If a request times out, the application should resolve the original payment's status before authorizing another one, and use duplicate-payment protection, such as a unique ID for each payment attempt, to reduce the risk of duplicate charges on retry. Delivery needs its own check, too: a transaction receipt confirms that money moved, but not that the service was delivered.

Which protocols support AI agent payments?

Open protocols can carry the payment exchange described above. With x402, a service responds to a request with HTTP 402, the web's "Payment Required" status code, along with its payment requirements. A compatible client then replies with a signed payment payload, a message that authorizes the payment. Verification, service execution, and settlement follow the selected payment scheme.

The broader ecosystem is standardizing other parts of this process. Emerging standards such as the Agentic Commerce Protocol and the Universal Commerce Protocol address commerce interactions, while the Model Context Protocol (MCP) connects AI applications to tools and data. Each of these interfaces still needs an underlying funding and authorization system: the wallets, stablecoins, spending policies, and session keys covered above.

Building accountable AI agent payments

For builders, the AI payments opportunity is concrete: give agents access to useful services under financial boundaries that operators can inspect and revoke. The essential pieces are the ones our coding agent relied on: a funded balance, protected signing credentials, enforceable spending rules, and reliable settlement. When appropriately configured and monitored, agents can acquire what a task needs and account for what it costs, all under human oversight and clearly defined security policies.

‍

‍

Arc is an open L1 blockchain launched by Arc Network Services LLC ("Arc LLC") and operated by a permissioned validator set. Arc LLC provides software services only and does not offer regulated financial or advisory services. Arc has not been reviewed or approved by the New York State Department of Financial Services or any other regulatory authority.

The Arc network is provided "as is" and "as available." Use of Arc involves inherent risks associated with blockchain technology, including smart contract vulnerabilities, network disruptions, and the absence of recourse for transaction errors or losses. The ability to transact on Arc depends on the ability to obtain and use USDC to pay gas fees. Neither Arc LLC nor any permissioned validator is responsible for the content, accuracy, legality, or functionality of third-party applications, protocols, or services built on or integrated with Arc. You are solely responsible for features or services you provide to users, including obtaining any necessary licenses or approvals and otherwise complying with applicable laws.

All Arc features may be modified, delayed, or cancelled at any time without notice. Nothing herein constitutes a commitment, warranty, guarantee or legal, regulatory, tax, or investment advice.

Agent Wallets, Agent Marketplace, Circle CLI, Nanopayments, and Circle Skills (together, the “Agent Stack”) are offered by Circle Technology Services, LLC (“CTS”). CTS is a software provider and does not provide financial, advisory, or marketplace services. Agent Stack enables users, developers, and their artificial intelligence agents ("AI Agents") to build, deploy, discover, hold and transact their own funds, and interact with other third-party AI Agents, applications and services. Transactions initiated by AI Agents are executed based on user-defined permissions and may occur without real-time human review. Circle does not verify, screen, monitor, or make any representations or warranties regarding the identity, legitimacy, safety, security, reliability, or performance of any AI Agent, including any AI Agent developed by you, other users, or third parties, or any AI Agent that interacts with the Agent Stack. AI Agents may operate unpredictably, may be controlled by unknown or unauthorized parties, and may expose users to security, financial, or other risks. Users are solely responsible for their use of these tools, reviewing commands before execution, and for evaluating associated risks. For additional details, please see the Agent Platform Terms of Use available at agents.circle.com/terms-of-use and the Circle Developer Terms of Service available at console.circle.com/legal/developer-terms.

Related posts

Navigate what’s next in the new internet financial system

What Is Agentic Commerce?

What Is Agentic Commerce?

November 4, 2026
What Is the Agentic Economy?

What Is the Agentic Economy?

October 27, 2026
Blog
How AI Agents Hold and Spend Money
how-ai-agents-hold-and-spend-money
November 19, 2026
Discover how AI agents hold and spend money using stablecoins, AI agent wallets, spending limits, delegated permissions, and session keys.
Agentic Economy
Agentic Economy