Ken Fromm

The recent discourse on agentic commerce on x led by @vigstar1 is really nice (instead of just funding announcements). But there’s a lot of it, so I asked @bot to grab relevant posts from @scottew @vigstar1 @alfongj @RonakDaya @regulatorynerd @soupsranjan @sophgoldb @amandapey @davistrazza @mttjon @heysamir_ and @noyesclt and turn it into a podcast episode: open.spotify.com/episode/5Sj… Summary: Core problem: when an AI agent buys something for you, the payment can go through and still be wrong. You asked for a refundable hotel and it booked the nonrefundable one. The card worked fine. Nobody in the chain can see what you actually asked for. Basics: a normal card purchase has five players. You, your bank, the network (Visa, Mastercard), the merchant, and the merchant’s processor. The merchant pays about 2 to 3% per sale, mostly to your bank. That fee funds your rewards, and it’s what everyone is fighting over. Agents pay in roughly four ways today: 1. Some type your real card into checkout, which works but is a security mess. 2. The networks built “agentic tokens,” which are the cleanest idea, but big banks like Chase don’t support them yet. 3. Others give the agent its own card. 4. And the newest option is a prepaid balance. The third option is what happened to @scottew. He picked his Amex in Muse. Link charged his Amex, then quietly created its own Visa to pay Best Buy. One item failed, Best Buy asked for that Visa’s security code, and he had no way to find it or know who to call. His ask is simple. Use the card I picked and let the store charge me. Swapping in a different card breaks rewards, muddies purchase protection, and leaves no clear place to get help. @RonakDaya’s guess on why it happened is that Best Buy likely isn’t set up for Link, and Amex can’t be tokenized for merchants like that yet, so Link fell back to a one-time card. The card you pick isn’t always the card that pays. So why would Stripe sit in the middle (@mttjon asked)? By putting its own card in the middle, Stripe earns the fee on the second leg. But the margin is thin, about $0.40 on $100 by @vigstar1’s estimate, and Visa and Mastercard dislike the setup. Hence the recent Link Balance launch: You preload money and the agent spends it without asking each time. Under the hood it’s a stablecoin plus a prepaid Visa from a small bank that’s exempt from debit fee caps, so it’s cheap to load and profitable when spent. The catch is that Link’s terms say agent mistakes like hallucinations and misreads don’t count as unauthorized. If the agent buys the wrong shoes, that’s on you. The balance also isn’t FDIC insured. The takes are split. @regulatorynerd thinks the card-in-the-middle approach is bad for consumers and the winner will tokenize your real card, and that agents built on janky checkouts will “wear the stink.” @krunalbhayani calls it Stripe chasing fees at everyone else’s expense. @vigstar1 says agentic tokens aren’t reliable yet, and the real issue is that nobody carries the risk when an agent gets your intent wrong. @noyesclt says forget “know your agent.” Banks want to know what you allowed and that the agent stayed inside it. The bigger picture, via @alfongj, is that Stripe has the merchants and now holds consumer money on its own rails. If both sides are on Stripe, a purchase could skip Visa entirely. @RonakDaya also zooms out past shopping. For personal purchases, a balance works because the money left and the permission left are the same number. If $637 is left, the agent can spend at most $637. Business use breaks that. “Renew approved vendors, max $10k each, $100k a month, no contracts over a year” is a set of rules, not a balance. Someone has to track what’s spent, reserved and still allowed across every agent, and he hasn’t seen anyone own that yet. His sharpest point is about mistakes that stay inside the rules. The agent books a $550 hotel under a $600 cap, and it’s awful. Every control worked. He sees it as a bad employee call, so whoever delegated owns it unless someone explicitly takes on that risk.