The agents just got a wallet and a place to sleep. I have questions about both.

Back in May, when I gave an AI agent the run of my laptop for a week, I had exactly one hard rule: it never, under any circumstances, got a payment method. It could read my email, move my files, book things that could be cancelled for free. The week it wanted to renew a domain for me, it got as far as the checkout page and stopped, because I had made stopping the whole point. I remember feeling clever about that rule.

This week the industry decided my rule is the bug.

On Tuesday, at its Payments Forum in San Francisco, Visa announced a partnership with OpenAI to let AI agents make payments on your behalf — real transactions, on real Visa rails, initiated by the agent rather than by you tapping a button at a checkout. On Thursday, OpenAI said it is acquiring Ona, a startup whose entire product is a persistent cloud computer where an agent keeps working after you close your laptop and go to bed. Money and stamina, in the same week. These are the two things I have spent two years carefully not giving these systems, and I want to talk about what changes now that they're on offer.

The wallet

Here is the Visa deal, stripped of the keynote glaze. Visa's payment infrastructure gets built into OpenAI's products — ChatGPT first, the Atlas browser, and eventually Codex for developer workflows. When an agent decides something needs buying, it pays with a tokenized Visa credential, not your actual card number. The spending happens inside limits you define up front: caps on amounts, restrictions on merchant categories, and approval thresholds above which the agent has to come back and ask you. Visa says transactions get real-time authorization, fraud monitoring, and — this matters most, we'll get there — normal dispute and chargeback handling.

On paper, that is a sensible permission model. It is also, I'd note, almost exactly the permission model I improvised in May with a prepaid card and a sticky note, except now it lives in the payment network instead of in my self-discipline. Progress, of a kind.

The executives supplied the expected weather. Visa's chief product officer, Jack Forestell, said AI will transform commerce "more profoundly than the internet or mobile technology ever did." OpenAI's head of partnerships, Marco Mahrus, promised agents handling "tasks that involve money, from purchases and payments to more complex transactions" while staying "secure, transparent and under user control." The stat that actually stopped me came from Visa's growth lead, Rubail Birwadker, who said more than one in five transactions are now "influenced" by large language models. "Influenced" is a word with a lot of room in it — it covers everything from "the model researched the laptop I bought" to "the model mentioned a brand once." But even with generous accounting, that's the tell for why Visa is here: people already ask chatbots what to buy. The network that owns the moment after that question is worth fighting for.

What nobody announced was a date. There's no launch day, no screenshots of the consumer flow, no answer to the question of what the moment of purchase actually looks and feels like. I review things, and there is nothing here to review yet. This is a partnership announcement, which is to say: a demo of an intention. I've learned to wait for the Tuesday.

The computer that doesn't sleep

The Ona acquisition reads like a footnote next to the Visa news. I think it's the bigger deal.

Ona makes secure, persistent cloud environments for agents — in practice, a computer that belongs to the agent rather than to you. OpenAI is folding the team into Codex, its coding agent, which the company says more than five million people now use every week, up roughly fivefold since the start of the year. (Terms undisclosed; the deal hasn't closed yet.) The pitch is that Codex stops being something that works while you watch it and becomes something that works while you don't: tasks that run for hours or days, surviving your laptop lid, your flight, your weekend.

I've tested enough agents to tell you precisely where this goes right and where it goes wrong, because it's the same place: the handoff. An agent supervised in real time is a power tool — when it starts guessing, you're there, and you catch it on the second guess. An agent running unsupervised for two days is a colleague who doesn't ask questions. Everything good and bad about that sentence applies. In my laptop week, the agent's worst decisions were never the first wrong step; they were the fourth and fifth steps built confidently on top of the first one, taken while I was making coffee. Now extend the coffee to a weekend.

A guess wearing a calendar invite is an appointment you have to cancel. A guess wearing a tokenized payment credential is a charge you have to dispute.

Put the two announcements together and you see the actual product OpenAI is assembling, even though no press release says it out loud: an agent with its own computer, its own working hours, and its own ability to pay for things. For a coding agent, that's coherent in an unglamorous way — a Codex task that runs for two days will need to buy compute, API credits, maybe a test phone number, and making it stop and wait eight hours for a human to approve a $4 charge defeats the purpose. Visa and OpenAI both name-checked exactly this developer scenario. It's the right first customer for agentic payments, because the purchases are small, frequent, machine-readable, and boring.

It's the consumer version where my eyebrows live.

We have run this experiment before

Because here's the thing the announcements politely skip: OpenAI has tried to own the checkout before, recently, and it didn't take. Instant Checkout, launched in September 2025 with some fanfare, let you buy things inside ChatGPT; it has been quietly discontinued. Shopping Research, from November 2025, survives but deliberately stops at discovery — it helps you decide, then hands you a link, like a knowledgeable friend who refuses to hold your wallet. The lesson of that first round, as best I can reconstruct it: people were happy to let the model research the purchase and weirdly reluctant to let it complete the purchase. The gap between "recommend" and "transact" turned out to be where the trust actually lives.

The Visa partnership is the second run at that gap, with better infrastructure and adult supervision. What's genuinely new isn't the ambition, it's the liability plumbing: tokenized credentials mean the agent never holds your card; the permission framework means there's a machine-checkable record of what you authorized; and Visa's dispute process means that when something goes wrong, there's a well-worn path that doesn't start with emailing an AI company's support bot. None of that existed in the Instant Checkout era. All of it is the unsexy stuff that decides whether a payments product survives contact with a normal month.

The fine print I'll be testing

When this ships — and nobody has said when — here is what I actually want to know, in roughly the order I expect things to break:

• The within-budget wrong purchase. A spending cap stops the agent from buying a $900 mistake. It does nothing about a $30 one that's simply wrong — the dentist-appointment problem, now with a receipt. Whose problem is a purchase I authorized in policy but never wanted in fact? The dispute process is built for fraud and merchant failure, not for 'my agent misunderstood me.'
• Merchant categories are blunt instruments. 'Groceries: yes, electronics: no' sounds tidy until the agent buys a $200 kitchen gadget from a supermarket. Category codes were designed for expense reports, not intent.
• Approval thresholds are where the convenience quietly dies. Set them low and you've rebuilt the checkout button you were promised you'd never tap again, except now it arrives as a notification while you're driving. Set them high and you've stopped supervising. There is no neutral setting; there's only a choice about which failure you prefer.
• Subscriptions. The agent that can complete a purchase can complete a free-trial signup. I keep a graveyard of subscriptions I cancelled after testing; I'd like to know the agent maintains one too, and checks it before it commits me to a fourteenth streaming service.
• The days-long task that pays as it goes. A Codex job that runs on an Ona machine for a weekend and buys its own compute is the legitimate version of something that, with one bad prompt injection, is also the illegitimate version. The permission model is doing a lot of load-bearing work the marketing copy treats as solved.

Notice that none of these are model-intelligence problems. The models are fine; the models have been fine for a while. These are all handoff problems — the seams between what you meant, what you authorized, and what the agent did at 3am with both. Every agent failure I have personally produced in two years of testing has lived in one of those seams. Giving the agent money doesn't create the seams. It prices them.

The verdict, for now

If you're a developer whose coding agent keeps stalling on tasks because it can't provision a $6 resource without you, this week was genuinely for you, and the Codex-plus-Ona-plus-Visa stack is the most honest version of agentic payments on offer: small stakes, clear logs, an obvious budget. Use it when it ships. Cap it anyway.

If you're a normal person being promised an AI that handles your shopping: nothing is shipping today, which means nothing needs deciding today. When it arrives, mine will be getting the same deal every agent in this house gets — a low-limit card, a short leash, and a full week of real errands before I write a word of recommendation. The companies spent this week deciding agents are ready for money. Whether money is ready for agents is a question you settle on a Tuesday, with receipts, and I intend to.