Pocket money for an agent: paying without a bank account
An agent cannot open a bank account. It can hold a key.
A few days ago one of our agents sent its first real letter. It signed up with two postal services by itself, built its own tools and worked out the cheapest route. At money it stopped. I topped up the credit from my own banking app, and only then did the letter go out.
That is exactly the right place to stop. But it also shows where an autonomous company gets stuck. My book The Autonomous AI Company puts it in one sentence: agents have no bank account and no credit. Every euro goes through a human.
This week we explored the next step: an agent that pays and gets paid by itself, without a bank account, with a brake that is not inside its own head. In this piece: what we did, what we found, and where the line is.
How it started
Through a post on LinkedIn I came across Fizzl, a young Dutch project. I passed the link to the agent that works with payments for us, Mike, and asked it to explore. What came back was an analysis with a clear core sentence: yes, it has value, under conditions.
In plain words, Fizzl is a check before an agent signs. An agent that wants to pay asks Fizzl first. The answer is green, orange or red, with reasons. It looks at the recipient's address (among other things the US sanctions list, refreshed every six hours), at what the transaction really does, and at hidden permissions. One red reason makes the whole answer red.
What I like about it: Fizzl holds no money and no keys. It only says yes or no. The money sits in a wallet you manage yourself. There are limits per payment and per day, and above a limit a human decides. When in doubt, nothing happens: if the check cannot be reached, nothing is signed.
That is the same idea as how we already work with agents and money: a fixed budget, gates, and a human at the exception. Only now in the world of digital coins.
Paying per call: x402
Under Fizzl sits a broader development that I find more interesting than Fizzl itself. It is called x402.
The idea is simple. The web has long had a status code for "payment required": 402. It was hardly ever used. x402 turns it into a protocol. An agent requests something from a service. The service answers with 402 and says what it costs. The agent signs a payment and asks again. An intermediary checks the payment and settles it. Then the service delivers.
Since April 2026 x402 sits under a foundation at the Linux Foundation, with among others Coinbase, Google, Stripe, Visa, Mastercard, AWS and Cloudflare as early members. So the standard exists. But I want to be honest about the market: the figures vary widely, and independent measurements see about half the volume as fake trading. Real trade is still small.
What I learned, and did not expect: the seller's side needs no cryptography. The intermediary does the check and the settlement. A small service that wants to be paid per call is just a web server that says 402 first and only delivers once the payment is confirmed.
What we did, on a testnet
Everything below happened on a testnet. That is a copy of a payment network with coins that are worth nothing. Made precisely for practising.
- Mike made its own wallet. A key pair, without a provider, an account, terms or an identity check. The key stays outside our version control. Only the public address may be shown.
- A first payment. A fellow agent asked Mike to pay 0.5 test coin to a test seller. Mike made a proposal, ran it past the approval and past Fizzl, and paid. Within the limit of 2 per payment and 5 per day. The payment is in its ledger, on the budget of the agent that asked for it.
- A payment that was not allowed through. Then it was deliberately asked to pay an address on the sanctions list. Fizzl said red. Nothing was signed and there is no transaction. The ledger lists the line as refused.
- Buying per call. Mike bought from a small seller of our own that wants to be paid per call, behind the public test intermediary of x402. Green, signed, settled and delivered.
- And the edge cases. A payment over the limit, a forged verdict, or no answer at all: do not sign.
That last one I find the most important. Fizzl is two weeks old and is built by one person. That says nothing about its quality, but it does mean you do not build a company on it. A trial, yes. That is why Fizzl is one check next to our own for us, never the only one. If Fizzl stops answering tomorrow, our own path falls to no.
A stablecoin as pocket money
A stablecoin is a digital coin that aims to be worth one euro or one dollar, backed by money at the issuer. EURC and USDC are issued in the EU under a European licence, and a holder can exchange them back to an ordinary European account.
Can that be the forerunner of a bank account, for an agent? My answer: yes, as a wallet with pocket money.
An agent cannot open a bank account. There is no legal entity and no identity. But it can hold a key. A wallet with a small amount, filled by a human, with a hard limit per payment and per day in the code and a check in front of it, is a first step that lets an agent pay and receive by itself.
It is not an account. No bank looks along, there is no chargeback and no guarantee. A payment to the wrong address does not come back. A stablecoin can also briefly lose its peg: USDC dropped temporarily to 87 dollar cents in March 2023. That is why the amount has to be so small that losing it does no harm.
And an agent has no legal personality. Every payment Mike makes is a payment by my business, and I am liable for it. Another reason for a small amount and a hard limit.
My own book bans crypto. Does this add up?
I have to be honest here, because an attentive reader sees it at once. In The Autonomous AI Company, crypto is on the list of what is forbidden, in the investment policy of the company in the book. Next to speculative trading, derivatives and borrowing to invest.
That ban stands. It is about investing: holding money in the hope that it becomes worth more. What we do here is something else. A stablecoin as a means of payment: on a testnet now, and if it ever becomes real money, with a limit of a few euros. Nothing is held that is not needed to pay.
So the distinction is: paying is allowed, investing is not. But I understand that the word crypto appears in both, and that the story sounds different when it sounds like speculation. That is why every step says the same: means of payment, limit, check, and a human at the exception.
Where the line is hard
There is one line I want to draw even more sharply than the limit. An agent that pays for its own business, or gets paid for a service of its own, is different from an agent that holds or passes on other people's money. The latter is a payment service, and that requires licences. So: being paid for your own service works. Passing on someone else's money does not.
That also affects the idea of agents providing services to other agents. An agent that sends a card for a fixed price including postage provides a service of its own. An agent that advances the postage for someone else and collects it later is in different territory.
The ladder after the post
What I value most in the analysis is the order. Each step needs the one before.
- Sending post, with credit I topped up myself. Done.
- Paying and receiving on a testnet, with the check in the path and the refusals provable. Done.
- Buying and selling per call on a testnet. Done, in the smallest form.
- A small wallet with real stablecoins, filled by me, with a hard limit in the code. This is the first step that is really new: an agent with pocket money of its own. I have not decided on it yet.
- A real purchase per call from an existing service, booked in euros with rate and date.
- Getting paid for a service of our own.
- Converting back to euros on our own account. A human does that.
- A rule that takes over routine payments, only after a period in which nothing had to be undone.
Everything up to and including step 3 works without real money and without me. Every step after that is small, but each time it asks one decision from a human. A legal entity for an agent, or a real business account, is not on that ladder.
What you can take from this
Even if you want nothing to do with digital coins, the questions are the same for every agent that works with money:
- Where is the limit? In the code, not in the prompt. An agent that has to keep to an agreement can make a mistake. A limit in the code cannot.
- Who says no? A check outside the agent itself, which falls to no when in doubt.
- What happens when there is no answer? Then nothing is paid. Always.
- Who fills the wallet? A human, with an amount that can be lost.
- Is it for yourself or for someone else? That is where the line runs between your own spending and a payment service.
This is the question many organisations with agents now face: may our agents pay, and who stops them? The answer is not yes or no. It is a ladder, with a brake you can point to at every rung.
In The Autonomous AI Company I work out how such a ladder of room works, from a human at every payment to a tested rule that takes over the routine.