Agentic wallets: when your money gets an API

An AI agent can book my flights, fill my cart and negotiate a refund. Right up until the moment money has to move. Then it stops and asks me to type my card number, because the one thing it cannot safely do is spend.

That gap is closing fast. The race right now is to give agents a wallet, a way to hold funds and pay on your behalf, with enough guardrails that you would actually let them. As a payments person, this is the most interesting thing happening in the industry, so I spent some time mapping it.

What an agentic wallet actually is

Strip away the hype and an agentic wallet is three things bolted together. An identity, so the world knows which agent is acting and whose money it is. A set of permissions, so the agent can spend fifty dollars on groceries but not wire your savings to a stranger. And a rail, the actual pipe the money travels down. Miss any one of those and you do not have a product, you have a liability.

The key word is scoped. A good agentic wallet never hands the agent a blank cheque. It hands over a narrow, revocable slice of your spending authority, the way you might give an assistant a company card with a limit and a category lock.

Why this is happening now

Two things had to be true. The models had to get good enough to act, not just chat. That happened. And the payments industry had to decide agents were worth building for, instead of something to block as fraud. That is happening now, quickly, because everyone can see where commerce is heading.

Here is the uncomfortable part for my industry. Every fraud system ever built assumes the buyer is a human with a heartbeat and a phone. An agent is neither. So the same networks that spent twenty years learning to tell humans from bots now have to learn to welcome the right bots and price them fairly.

The market, mapped

Two worlds are converging on the same problem from opposite ends.

The card networks are extending what they already own. Visa's Intelligent Commerce and Mastercard's Agent Pay both do a version of the same thing: issue an agent its own credential, tied to your account, with rules attached. This is the safe, familiar path. It keeps agents inside rails that already reach every merchant on earth.

The platforms are making checkout agent-ready. Stripe shipped tooling so an agent can pay through the same infrastructure that already runs a huge slice of online commerce, and the checkout flows appearing inside chat assistants are early versions of the same idea. The pitch is simple: the merchant does almost nothing, the agent just works.

Then there is the crypto-native camp, and this is where it gets genuinely new. Coinbase's x402 revives an old, unused corner of the web, the HTTP 402 Payment Required status code, and turns it into a way for software to pay software. An agent hits an API, gets told it costs a fraction of a cent, pays in stablecoins, and gets its answer, all in one round trip. No account, no signup, no card. For machine-to-machine payments that is a real unlock.

And sitting across all of it are the protocols trying to be neutral. Google's AP2 is the most ambitious: a shared language for an agent to prove what it is allowed to do, that the card networks, the crypto players and the platforms have all put their names to. Whether one protocol wins or we end up with five is the open question.

Two rails, and why agents will use both

The lazy take is card rails versus crypto rails, pick a side. The useful take is that they are good at different jobs, and an agent does not care about your tribe.

If the agent is buying a real thing, a hotel night, a pair of shoes, a software subscription, card rails win. They reach everywhere, and crucially they are reversible. When your agent books the wrong hotel, you want a chargeback, not a lesson about the finality of blockchains.

If the agent is buying compute, data, an API call, something that costs a hundredth of a cent and happens ten thousand times an hour, card rails fall apart. The fees alone make it impossible. That is exactly the job stablecoin rails were built for. Tiny, instant, final, machine to machine.

So the real design question is not which rail. It is how you let one agent, under one set of your rules, reach for whichever rail fits the task. That is the problem AP2 and the wallet layer are actually trying to solve.

The hard parts nobody has fully solved

It is easy to demo an agent buying a coffee. The hard parts show up right after.

Authorization is the whole game. "Spend up to fifty dollars on groceries this week" sounds simple until the agent finds a great deal on something that is technically groceries and technically not what you meant. Mandates and limits are a start. Intent is much harder.

Then there is the dispute problem. Chargebacks were designed around a human who can say they did not authorize a charge. When an agent you authorized makes a purchase you regret, who is liable? You? The agent's developer? The merchant who happily sold to a bot? Nobody has a clean answer, and the answer will shape the whole market.

And reconciliation, the unglamorous one I always come back to. When thousands of tiny agent payments hit your books across two or three different rails, the finance team still has to tie every one back to a decision a human can defend. Agentic payments do not remove that job. They multiply it.

Where I think this goes

A few predictions, held loosely.

Mandates become the primitive. The important object stops being the card and becomes the permission: a signed, scoped, revocable grant that says what an agent may do with your money. Whoever owns that layer owns the relationship.

Agents become a customer segment. Merchants will start optimizing for agent buyers the way they once optimized for mobile. Clean APIs, machine-readable prices, agent-friendly checkout. The sites that make it easy will win a new kind of traffic.

Micropayments finally happen, for machines not people. We spent twenty years failing to make humans pay a penny per article. Agents will happily pay a penny per API call, all day, because they do not feel the friction we do. That is the market x402 is chasing.

Human-in-the-loop shrinks, but never to zero. Low value and reversible, let the agent run. High value or irreversible, a human still taps approve. The interesting work is drawing that line well, per person, and moving it as trust builds.

The wallet is the interface

The wallet is quietly becoming the most important interface in AI. Not the chat window, the wallet. Because the moment an agent can pay, it stops being a clever demo and starts being an economic actor. That is a bigger shift than it looks, and most of the plumbing for it is being poured right now.

I do not know which of these companies wins. I am fairly sure the pattern does: scoped authority, the right rail for the job and a receipt a human can check. Everything else is detail.