> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Agentic card tokens went from zero to millions weekly
- URL: https://nnutshell.com/agentic-card-tokens-went-from-zero-to-millions-weekly/
- Published: 2026-10-22T11:37:49.000Z
- Updated: 2026-10-24T03:10:26.000Z
- Description: Terry Angelos, chief strategy officer of VGS, talks with Andreessen Horowitz's Alex Rampell about agents getting payment credentials, the wallet wars, and unresolved liability when an AI makes a purchase.
- Author: Admin
- Tags: Terry Angelos, Alex Rampell, VGS, Stripe, Amazon, Walmart, Payments, Agentic AI, Fintech, E-commerce, #rec-2026-10-15, #Import 2026-10-25 09:54

*Recorded 15 October 2026 · 50 min*

Agents stopped being spectators and became actors in the last 3 months, and the reason is payment credentials. Once an agent can hold a card, it moves from recommending flowers to buying them, and every wallet is racing to connect to it.

Terry Angelos is chief strategy officer and board member at VGS, a payment vault that secures credentials for about a third of e-commerce GMV and holds some 10 billion of them. Alex Rampell of Andreessen Horowitz led the VGS investment and co-founded TrialPay with Angelos in 2006\. He interviews him here.

The tension is how the credential gets handed over. A walled-garden workaround functions but breaks chargebacks and merchant relationships. Card-network agentic tokens fix that. Yet nobody has settled who pays when an agent gets it wrong, and Amazon has already chosen to block agents.

1. **Tokens replace card numbers** — [▶ 4:51](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=291s&ref=nnutshell.com)  
Merchants no longer store and decrypt 16-digit PANs. They request a network token from Visa, Mastercard, Amex or Discover that is bound to their merchant domain, so a breach yields a number that works nowhere else. Churn makes this hard: about a third of cards turn over each year, 25% from the 4-year expiry and 5% from events like Capital One moving cards from Visa to Discover. A $1 billion merchant runs 3-5 processors, so vaulting and updates multiply. VGS sits as a neutral vault across all of them.
2. **Outsourcing PCI scope** — [▶ 10:29](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=629s&ref=nnutshell.com)  
An AI chat app asked to buy flowers does not want to store a card, and even if it does, it must pass the card on to the merchant. A third party that already knows how to do this is the better option. PCI DSS runs over 100 pages, and every bank demands compliance: no processing or issuing without it. VGS takes PCI out of scope, so a company can "just run my business like a normal business." That is why a lot of AI companies have chosen VGS.
3. **Spectators become actors** — [▶ 11:56](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=716s&ref=nnutshell.com)  
In the last 3 months agents moved from spectators to actors because "they effectively have arms and legs, which are payment credentials." Angelos asked ChatGPT about a piano teacher for his 7-year-old a while ago and got recommendations. Last night an agent found teachers, asked questions, checked a schedule that includes soccer 3 days a week, and booked the first lesson. That is "infinitely more useful" than dumping another task on the user. Utility jumps once an agent can transact.
4. **Wallet wars scoreboard** — [▶ 13:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=823s&ref=nnutshell.com)  
Agents want wallets with the most consumers holding credentials and the broadest merchant reach. Stripe Link has about 300 million consumers and about 1 million merchants. Shop Pay has a similar consumer base and about 1.7 million merchants. PayPal has over 400 million accounts and about 30 million merchants. The earlier fight was top of wallet: issuers like Chase and Amex paid tens of millions to Uber and Airbnb to be the default card. The new fight adds a layer above the card, wallets that process the transaction.
5. **Stripe, the AOL** — [▶ 20:05](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1205s&ref=nnutshell.com)  
Angelos asked an agent for outdoor lights. It found them at Walmart, where Link has no connection. Stripe issued a one-time Visa virtual card, bought the lights, then charged him in a second transaction. It worked, but his bank would see only a Stripe charge. A tester who called Amex about a dispute was given a number that rang a Stripe engineer's cell phone. Walmart's $6.99 delivery upsell to Walmart Plus failed on the one-time card. He calls Stripe "the AOL of agentic commerce", as a compliment.
6. **Agentic tokens at scale** — [▶ 24:04](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1444s&ref=nnutshell.com)  
The card networks' new agentic tokens let a wallet hand a credential to an agent, which passes it to the merchant. The agent declares it is verified and will spend no more than $100\. The token is just a new 16-digit number, and the customer's real card is charged, so chargeback rules and follow-on sales like Walmart Plus survive. Three months ago volume was zero. VGS, by far the largest provisioner, now does a few million tokens per week for agents, wallets and some merchants. Angelos expects the shift from walled garden to open standards.
7. **Vaults go multiplayer** — [▶ 28:09](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1689s&ref=nnutshell.com)  
VGS holds about 10 billion credentials, a role Rampell calls very boring: retrieve, charge, never touch the card again. Angelos says vaults are going from single player to multiplayer. Multiple processors already make the vault serve Chase, Stripe, Adyen, Braintree and Razorpay. Agents add another party: a frontier lab that vaulted cards for subscriptions can now give the same credential to an agent under open standards. "Any business that is going from single player to multiplayer inherently becomes more interesting." He cites 1Password as another example.
8. **Liability gap unresolved** — [▶ 32:18](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1938s&ref=nnutshell.com)  
A transaction needs three checks: is the card real, is the person its owner, is the person of sound mind. Networks now handle the first two. Agents must register with Visa, Mastercard or Amex, and the token carries context, like Angelos's San Diego hotel under $500\. The third, whether the agent hallucinated dates, is not understood yet, and the dispute mechanism is unresolved. Merchants will say delegated authority leaves no excuse, as in the friendly-fraud era. Angelos thinks agents will probably have to take on some responsibility.
9. **Merchants split on agents** — [▶ 35:31](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2131s&ref=nnutshell.com)  
Amazon recently blocked an agent. Rampell's read: it is an advertising company, eyeballs are its highest-value SKU, and an agent has none, so antagonism is rational, not Luddite. Walmart has not blocked and may gain. Shopify's small merchants have no choice. Angelos sees fear at Amazon, greed at Walmart, necessity for SMBs. Open banking is the precedent: Plaid and Chase settled on access only through the bank's tokenizing, possibly paid APIs. Merchants may field their own agents, like a Best Buy salesperson. He thinks consumers probably win, but does not know how.
10. **Over- and underhyped** — [▶ 41:01](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2461s&ref=nnutshell.com)  
Overhyped: a bank run as agents chase savings yield. Angelos keeps savings in Mercury for the better rate but pays bills from Chase because bill pay just works better, and that friction persists. Underhyped: unmonetized demand. His cycling group chat's friend Seth recommends new tires every 6 months. An agent in that chat would buy each time, so Rampell is dramatically under-monetized on bike gear. Agents will also cancel forgotten subscriptions. Human desires, he says, are infinite, and the always-on nature of agents has yet to be felt.
11. **Asymmetry and broken rewards** — [▶ 43:40](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2620s&ref=nnutshell.com)  
Early agent wins come from asymmetry: few people have the tool and counterparties don't know. Rampell says the 7:00 a.m. reservation call can't scale once everyone has one, and he is waiting for Comcast to demand a human. Angelos adds that rewards depend on breakage: a big dining reward is funded because customers also spend on fuel. If agents only use whichever card earns points, the whole points ecosystem falls apart. New terms will follow, which he says could be better or worse.

How an agent gets a card matters more than the fact that it has one. A card number is not just a number. It carries a set of rules: who can dispute a charge, who owns the relationship with the customer, who eats the loss when something goes wrong. Hand the number over badly and the rules stay behind. The first agent-commerce plumbing shows exactly how that happens.

Terry Angelos hit the problem himself. He asked an agent to find outdoor lights for an event at his house. It found them at Walmart and loaded the cart. Stripe Link held his card, but Link reaches only about 1 million merchants and Walmart was not one of them. So Stripe issued its own one-time Visa virtual card, bought the lights, then charged Angelos in a second transaction. The lights arrived and the amount was right. Then it got awkward. Walmart emailed an offer to join Walmart Plus and credit the $6.99 delivery fee. The click failed, because the card Walmart held was the one-time card. Walmart pays to acquire customers and expects to upsell them. The workaround broke that. It also broke disputes. Capital One would see a Stripe charge, not a Walmart charge. One tester who called Amex was given a number that rang the cell phone of a Stripe engineer. Angelos calls Stripe "the AOL of agentic commerce" and means it kindly. It works. It is also a walled garden.

Chargebacks are the point. Visa and Mastercard intermediate everything and set the rules, so a buyer who pays for lights that never arrive has somewhere to complain and a neutral party to adjudicate. Alex Rampell contrasts crypto, a bearer instrument: pay and you are never getting it back. The trust mechanism helps merchants too, because shoppers buy more readily when they can get their money back. Clever engineering that routes around the networks gives that up.

The fix is the networks' own standard, agentic tokens. A wallet such as Link, Shop Pay or PayPal hands a credential to the agent. The agent declares that it is verified and that it will spend no more than $100\. The token is just a new 16-digit number, which the agent gives to Walmart. Walmart charges the customer's actual Capital One Visa. Chargeback rules hold. The follow-on Walmart Plus sale works. The adoption curve is steep. Three months ago these tokens did not exist. VGS now provisions a few million per week and says it is by far the largest provisioner, serving agents, wallets and some merchants. Angelos expects the market to move from walled garden to open standards.

That is why a boring vault turns interesting. The vault stops being single player. A merchant with $1 billion in GMV runs 3-5 processors, so one vault feeds Chase, Stripe, Adyen and Braintree. A frontier lab that vaulted cards to bill subscriptions can now let an agent use the same credential. VGS holds about 10 billion credentials, and a third of them need updating every year. In Angelos's words, "any business that is going from single player to multiplayer inherently becomes more interesting." Wallets feel the same pull. An agent wants a wallet with many consumers and broad merchant reach. Link has about 300 million consumers and 1 million merchants. Shop Pay has a similar consumer base and about 1.7 million merchants. PayPal has over 400 million accounts and about 30 million merchants. Link's 1 million is exactly why the Walmart workaround existed. The old wallet fight was over top of wallet: issuers paid tens of millions to Uber and Airbnb to be the default card. This one adds a layer above the card.

The obvious objection is that merchants will not play. Amazon recently blocked an agent. Rampell's reading is that Amazon is an advertising company, the highest-value SKU is the eyeballs it sells, and an agent has none. Blocking is rational, not Luddite. Walmart has not blocked, and with the best last-mile delivery on Earth it may gain from agents. Shopify's small merchants have little choice, since agents are a vital new channel. Fear at Amazon, greed at Walmart, necessity for the SMBs. Open banking is the precedent. Plaid and Chase ended in a negotiated settlement: access yes, but only through the bank's APIs, tokenized, possibly for a fee. Retail may produce the same, with merchants running their own opinionated agents, like a Best Buy salesperson working on commission. Rampell adds the counter-pressure: a merchant that inserts 15 captchas will lose transactions to one that welcomes the agent. Angelos thinks consumers win in the end, but admits he does not know what that looks like.

The harder problem is who pays when the agent is wrong. Rampell frames a transaction as three checks: the card is real, the person is its owner, and the person is of sound mind. The networks now handle the first two. Agents must register with Visa, Mastercard or Amex, and the token carries context. Angelos's example was a San Diego hotel for his daughter's gymnastics tournament, capped at $500\. The third check has no answer. If the agent gets the dates wrong, the network did everything right and the dispute mechanism is unresolved. Merchants will not care. Delegated authority means the cardholder authorised someone, human or AI. The friendly-fraud era shows the risk: a hypothetical $10,000 virtual-gems bill disowned at month end, and banks siding with consumers on small disputes, which created moral hazard. Angelos's answer is a hope that agents will take on some responsibility. Rampell imagines agents competing by absorbing it.

The early wins also rest on a temporary edge. Rampell notes that agents beat reservation lines and phone trees because few people have them and counterparties do not know. That does not scale. He is waiting for Comcast to demand a human. Angelos sees GDP upside instead: unmonetized demand, like buying the tires a friend recommends in a cycling group chat every time, and cancelling forgotten subscriptions. He dismisses a savings bank run as hype, since bill pay at Chase just works better than Mercury's higher rate. Anything built on breakage is exposed. Rewards pay a big headline rate on dining and profit when you also spend on fuel. If agents use whichever card earns points, that ecosystem falls apart. New terms will follow. Angelos says "maybe that's better, maybe that's worse."

**The open question.** Who eats the loss when an agent books the wrong dates? Angelos says the networks have done everything right once they validate the agent and the instruction, and the dispute mechanism for a mistaken purchase is unresolved. His answer is that agents will probably have to take on some responsibility, and Rampell floats agents competing by absorbing it. That is a prediction, not a mechanism. Merchants want no excuses once authority is delegated, and the friendly-fraud era shows how small disputes drift to consumers and create moral hazard. Watch which agent first offers to underwrite its own errors.

Watch the full video at [https://www.youtube.com/watch?v=q8cb1XL9sXA](https://www.youtube.com/watch?v=q8cb1XL9sXA&ref=nnutshell.com).