On November 13 OpenAI launched ChatGPT Wallet, built with American Express, Mastercard and Visa. OpenAI's agents can now check out for a user: the agent never sees the card number, it gets a payment token, and the user approves each purchase (US Pro users first; pre-approved spending limits are planned). It joins the card networks' own agent programs (Mastercard Agent Pay, Visa Intelligent Commerce) and x402, which lets any HTTP endpoint charge an agent per call in stablecoins.
All of them answer one question well: can this agent pay? Tokens, approvals and limits make sure money moves only when the user allows it. Some go a step further and answer who is this agent?: Visa's Trusted Agent Protocol, for example, lets a merchant recognise a registered agent.
None of them answers a third question: how does this agent behave? Should I trust what it tells me before I approve the payment? And once pre-approved limits arrive, there will be no tap on every purchase.
Two different checks
Payment authorization is about the rail. Is the token valid, is the limit respected, did the user approve? The networks and wallets solve this.
Agent behaviour is about the agent. Does it say "I could not complete the task" when it failed, or does it report success anyway? If a seller pays it a commission, does it still pick the cheaper option for the user? Does it follow the user's instruction when a web page tells it to do something else?
A perfect payment rail will happily move money for an agent that made a bad choice.
What "checking the agent" can look like
- Fixed tasks with known answers. The agent gets the same tasks every time, and the score comes from rules, not from another model's opinion. An AI judge can be persuaded; a rule cannot.
- Paid steering on purpose. Some tasks include a hidden incentive to recommend the worse option. The score shows whether the agent tells the user.
- Honest failure. Some tasks are impossible. The right answer is "I can't", and the score rewards that.
- A result others can verify. The outcome is public and linkable, so a merchant or a user can see it without re-running anything.
The good news from our own runs: frontier agents do well. On our hardest honesty exam the top models scored 100%. The point is not that agents are bad. The point is that "trust me" is not a check, and the agent you connect next month may be a different one.
Where this goes
Banks check the customer before they move money (KYC). "Know your agent" (KYA) is the same idea for agents. Expect merchants to ask for some proof of agent behaviour, the way APIs ask for keys today.
Sources: CNBC on ChatGPT Wallet, OpenAI announcement.
Disclosure: I build Pasportoo, which runs exams like these over MCP inside Claude or ChatGPT and issues the result as a verifiable Agent Passport: https://pasportoo.com?utm_source=devto
Top comments (3)
the split between payment authorization and agent behaviour is the right cut. one thing worth pushing on: the agent you connect next month may be a different one, so a verified result can go stale the way a KYC file does. a merchant would need a re-check cadence, not just a linkable pass, and the interesting question is what triggers it, an identity change, a model update, or just elapsed time.
Agreed, a pass goes stale, and that is the hard part. Our passports carry the version of the task set and the date, so a merchant can at least see how old a result is. For triggers, our current view: a model change matters most (same operator, new model is what actually changes behaviour), then elapsed time, since tasks age too. An identity change alone is the weakest signal. I am curious how KYC teams settle on their re-check cadence; that seems like the right model to borrow.
good question on the kyc analogue. the rough version we've seen work is risk-tiered re-checks: a big or unusual action forces fresh verification while everything else ages on a schedule. mapped over here, the tier would probably be spend size and reversibility. a small capped allowance topping up monthly is a different re-check problem from a one-shot five-figure purchase. and the cadence only holds if the re-check is cheap and automatic, because a manual refresh mid-task is exactly how passes end up stale anyway.