Do You Need a Token to Pay AI Agents?
No. Machine payments need a stable settlement asset, and a project token is the opposite of that. Native tokens are genuinely required for gas on the network you settle on, which is a small, bounded exposure. Using one as the unit of account turns every metered call into a currency position neither party wanted.
Two roles that get conflated
Native tokens do a specific job: they pay for computation and secure the network. If you transact on a chain, you need some of its token to pay fees. That is a real, unavoidable requirement and it is small, bounded, and operationally easy to manage.
Settlement is a different role. It is the answer to what an API call costs and what the seller receives. Nothing about paying network fees implies that the price of a service should be denominated in the same asset.
Most agent infrastructure projects that launch a token conflate these, usually because the token is also the business model. The pitch is that agents will transact in the token, which creates demand for it. That may be good for the token. It is worse for the agents, and worse for the sellers they are trying to buy from.
What breaks when the settlement asset moves
Quotes stop being quotes. In a payment flow where the server prices a call and the client pays moments later, a moving asset means the amount agreed and the amount transferred are different. Someone absorbs that difference on every transaction.
Agents cannot compare offers. An agent evaluating two providers priced in two different volatile assets is doing currency conversion rather than procurement, using price data that may itself be stale.
Sellers inherit a conversion problem. A provider receiving a project token has to hold it or convert it. Holding is a position they did not choose. Converting costs fees and slippage and takes operational work. Either way it makes accepting you more expensive than accepting a stable asset, and sellers optimize for their own convenience.
Accounting gets harder. Revenue in a volatile asset has to be marked, and the value at receipt differs from the value at conversion. That is a real reporting burden for a small provider selling metered calls.
Acquisition becomes a prerequisite. If your agent must first obtain a token to buy anything, you have added a step in front of a purchase that was supposed to be a single request. That step needs liquidity, a venue, and error handling, and it fails in ways your payment flow cannot recover from.
When a token requirement is legitimate
Gas on the network you settle on. Staking or bonding where an economic guarantee is genuinely being purchased, such as collateral behind a service commitment. Governance, if you actually want a vote. These are specific and bounded. Notice that none of them require the price of an API call to be denominated in the token.
How to keep exposure minimal
Hold a working balance of the native asset for fees, sized against expected transaction volume with a margin, and top it up automatically rather than reactively. An agent that cannot transact because it ran out of gas is a failure mode you will hit at the least convenient moment.
Settle in a stable asset. This is what makes a quoted price a real price and what makes your counterparties willing to accept payment without a conversion step.
Avoid designs where an agent must acquire an asset before it can transact. Every acquisition hop is a place the flow breaks, and it breaks in a way that leaves the agent unable to complete the task rather than gracefully declining it.
And treat token-gated access as a red flag when evaluating providers. If a service requires you to hold its token to use it, that requirement is doing something other than serving you, and it is an ongoing operational cost you carry for the privilege.
This is why CryptoCadet settles in USDC on Base, with no protocol token in the payment path. The agent holds a bounded USDC float authorized by session keys, custody stays in the OS keychain, and the only native asset exposure is the small balance needed for fees. Prices quoted are prices paid, and sellers receive something they can account for without conversion.
Evaluating an agent infrastructure project
Ask four questions and the answers arrive quickly.
What is the token actually required for, mechanically, in the flow you care about. If the answer is that it captures value from usage, that is a business model rather than a technical requirement.
Can you use the system while holding only a stable asset plus fees. If not, the token is a toll rather than a component.
What happens to a seller who receives it. If they immediately convert, the token added a step and removed nothing.
And what does the design assume about liquidity. Systems that work only when a market is deep and orderly are systems that fail during the periods when you most need them to work.
None of this means tokens are useless. It means that for the specific job of paying for a metered service, a stable unit of account is the requirement, and a volatile asset in that role is a cost passed to both sides of every transaction.
Frequently asked questions
- Do AI agents need a crypto token to make payments?
- They need the network's native asset for fees, which is a small bounded requirement. They do not need a project token as the settlement asset. Machine payments require a stable unit of account so that a quoted price is still the price at settlement, which a volatile token cannot provide.
- What is wrong with settling agent payments in a project token?
- Quotes stop matching payments as the price moves, agents comparing providers end up doing currency conversion instead of procurement, sellers inherit a position they did not choose or a conversion cost, accounting requires marking revenue, and acquiring the token becomes a prerequisite step that can fail.
- When is holding a native token actually necessary?
- For gas on the network you settle on, for staking or bonding where an economic guarantee is genuinely being purchased, and for governance if you want a vote. All three are specific and bounded, and none of them require the price of a service to be denominated in that token.
- How much native token should an agent hold?
- A working balance for fees, sized against expected transaction volume with margin, topped up automatically rather than when it runs out. An agent that cannot transact because it exhausted its gas balance fails at the least convenient moment, usually mid-task with a payment partially completed.