Control for agentic payments should start at infrastructure

Booking travel or paying for subscriptions or for running procurement through Claude or a custom GPT wrapper no confirm button is required anymore. The capability side is mostly solved. What doesn't get talked about enough is what happens when it goes sideways. A stored card sitting in the agent's context means it holds that access the whole session. One bad tool call and it's spending outside what you intended with nothing at the infrastructure level stopping it. Real time card issuance is the cleaner model. Agent requests a card for the specific transaction, purchase completes, card cancels and nothing persists. Who is running agent initiated payments in production right now and what does the architecture look like?