Context
This episode is about the payments layer behind agentic commerce. Terry Angelos explains why AI agents become far more useful once they can hold payment credentials, and why that requires new infrastructure around wallets, tokens, merchant acceptance, and liability. He speaks from VGS's vantage point as a company that secures payment credentials for about a third of e-commerce by GMV and is now provisioning millions of agentic tokens each week. For PMs, the episode matters because many AI shopping features will succeed or fail on hidden details like chargebacks, merchant incentives, and who still owns the customer relationship after an agent checks out.
The Big Idea
AI agents become real commerce actors only when they can spend safely, and safe spending is not the same thing as successful checkout.
The hard part is preserving everything around the payment, chargebacks, merchant relationships, follow-on purchases, and liability, while letting an agent act on a user's behalf.
The episode's core point is that payments infrastructure is what turns an agent from a recommendation engine into something that can actually complete work. But the winning systems will not be the ones that merely get a transaction through once. They will be the ones that keep the existing trust model of card payments intact.
Key Insights
Payment access changes everything
Terry's simplest framing is also the most important: agents moved from spectators to actors once they got access to payment credentials. Before that, an agent could recommend flowers, a piano teacher, or a hotel, but it still dumped the final work back on the user. Once the agent can pay, it can complete the task end to end, which makes the product feel dramatically more useful instead of merely helpful.
Neutral vaults gain leverage
What used to look like boring infrastructure becomes strategic in an agent world.
- A large merchant often uses 3 to 5 processors, not one, because costs, approval rates, and regional coverage differ.
- About a third of cards turn over each year through expiration, fraud, or events like a network change.
- A neutral vault like VGS stores the credential once, manages the network tokens around it, and lets the merchant route transactions without getting locked into one processor.
- That same neutral position now matters even more because the credential is no longer moving only between a merchant and a processor. It may now move among wallets, agents, merchants, banks, and networks.
For PMs, the lesson is that agentic commerce creates value for boring middle-layer infrastructure. The more parties that need safe access, the more powerful neutral brokering becomes.
Wallet scale becomes distribution
The episode argues that agentic commerce has restarted the wallet wars.
- An agent wants a wallet with lots of consumers who already have credentials stored.
- An agent also wants a wallet with broad merchant reach.
- Terry's examples:
- Stripe Link: about 300 million consumer accounts and roughly 1 million merchants.
- Shop Pay: similar consumer scale and about 1.7 million merchants.
- PayPal: over 400 million accounts and around 30 million merchants.
This matters because wallet success is not just a brand or UX story. It is a two-sided network story. A wallet becomes more valuable to agents when it already has both user permission and merchant acceptance.
Closed loops break trust
The episode's most concrete product lesson is that making the first purchase work is not enough.
Terry describes using an agent to buy outdoor lights from Walmart. Stripe Link held his credential, but Link did not have a direct relationship with Walmart. So Stripe bought the item on Terry's behalf using its own one-time virtual card and then charged Terry in a second transaction. The purchase worked, but the structure caused two problems:
- If Terry later needed help with a dispute, his bank would only see a Stripe transaction, not a Walmart transaction.
- Walmart could not complete a later Walmart Plus subscription flow, because the card it had on file was Stripe's temporary card, not Terry's real one.
For PMs, this is a strong warning. If your design hides the real customer payment relationship behind an intermediary, you may solve checkout while breaking refunds, upsells, subscriptions, customer support, and chargebacks.
Agentic tokens keep rails intact
The card networks' answer is a new standard, agentic tokens.
- A wallet can hand a token to an agent in a controlled way.
- The agent can hand that token to a merchant.
- The merchant can run the transaction as the user's actual card, rather than an intermediary's substitute card.
- That preserves existing card-network behavior, including follow-on merchant relationships and dispute processes.
Terry says this went from zero three months earlier to VGS provisioning a few million agentic tokens per week. The implication is important: the market is moving from clever workarounds to standard rails surprisingly fast.
Merchant incentives will split
Not every merchant should want the same future.
- Alex argues Amazon has a real reason to resist agents, because Amazon monetizes human attention through advertising and merchandising. An agent removes the eyeballs.
- Walmart may have more reason to welcome agents if inventory breadth and last-mile delivery matter more than preserving a browsing surface.
- Smaller merchants on Shopify may see agents as a new acquisition channel they cannot afford to ignore.
The PM takeaway is that agentic commerce is not a universal good or bad. Whether a merchant welcomes agents depends on where its profit comes from.
Agents attack economic breakage
The episode makes a subtle economic point that matters well beyond payments.
Agents may increase spending by surfacing latent demand. Terry and Alex use a cycling example: if an agent sat in a group chat and noticed repeated recommendations for better tires, it could convert those moments into real purchases that the user would otherwise forget.
At the same time, agents may reduce spending by eliminating waste:
- canceling forgotten subscriptions,
- avoiding trial-to-paid traps,
- choosing the best rewards card more consistently,
- and optimizing terms and conditions that consumers usually ignore.
That means AI can both create more GDP and destroy business models built on breakage, meaning revenue that depends on customer inattention.
Mental Models & Frameworks
Spectator to actor threshold
Use Terry's framing as a product test: an AI feature becomes materially more useful when it can act, not just recommend.
Ask three questions:
- Can the AI complete the job end to end?
- Does the user need to re-enter data or approve every minor step?
- Does the AI have the permissions, credentials, and rails needed to finish the transaction safely?
Use this model when evaluating whether an AI feature is truly automating work or just creating a more conversational to-do list.
Three checks for delegated spend
Alex breaks delegated payments into three checks.
- Card validity: is the payment credential itself real and usable?
- Authority validity: is the party making the purchase authorized to use that credential?
- Decision validity: did the agent actually make the right choice, such as the right hotel dates or item?
The first two problems are increasingly solvable through tokenization, registration, and network rules. The third remains much less settled. Use this model when reviewing any agentic checkout flow, because a system that only solves payment authorization has not solved the full user-trust problem.
AOL versus internet path
Terry compares current agentic commerce to the early internet.
A closed system can make things work quickly inside a controlled environment. That is the AOL phase. Stripe's workaround for unsupported merchants fits this pattern. It is convenient and gets early adoption.
But open standards are what let the broader ecosystem keep its normal behaviors. In payments, that means the merchant still knows the buyer, the bank still sees the merchant, and existing dispute rules still apply. Use this model when choosing between short-term product convenience and long-term ecosystem compatibility.
Single-player to multiplayer vault
A vault becomes more valuable when one stored credential can safely serve multiple parties under clear rules.
Before agentic commerce, a stored credential might mainly support a merchant's own recurring billing. Now the same stored credential may need to support:
- multiple processors,
- network token refreshes,
- agent-initiated purchases,
- wallet handoffs,
- and bank-side tokenization.
Terry's point is that infrastructure gets more interesting when it shifts from single-player to multiplayer. PMs can use this model to spot hidden platform opportunities inside systems that used to look like commodity back-end tools.
Decision Principles
Principle: Preserve existing protections
- When: building or evaluating any agent payment flow.
- Why: consumers and merchants trust card networks because they already handle identity checks, chargebacks, and follow-on servicing. A flow that bypasses those protections may boost conversion in the short run but creates a weaker product system.
Principle: Choose your agent stance
- When: deciding whether a merchant product should welcome, limit, or block AI agents.
- Why: the right choice depends on the business model. A merchant that profits from browsing and ads may resist agents, while a merchant that wants more completed orders may treat agents as a growth channel.
Principle: Optimize the second transaction
- When: designing a checkout experience that uses an intermediary wallet or agent.
- Why: the first authorization is only part of commerce. Refunds, subscriptions, upsells, loyalty flows, and support all depend on the merchant understanding who really paid and how.
Principle: Treat wallets as two-sided
- When: choosing which wallet, credential layer, or partner to integrate.
- Why: a wallet's value comes from both sides of the network, stored consumer credentials and merchant acceptance. Large user counts alone do not make a wallet useful to agents if merchants cannot accept it where the agent wants to shop.
Trade-offs & Nuance
Simplicity versus compatibility
A closed-loop workaround can make an agent purchase work today with little merchant integration. That is attractive for speed. But the cost is weaker compatibility with the rest of the payments system, especially disputes, recurring use, and merchant-side customer relationships.
Delegation versus liability
Giving an agent more autonomy reduces friction and makes the product feel magical. It also raises harder questions about who is liable when the card was valid and authorized but the agent chose the wrong thing. The episode is clear that this part is still unresolved.
Eyeballs versus transaction volume
Agent traffic may increase conversion and remove checkout friction. It may also eliminate high-margin surfaces like ads, cross-sells, and merchandising. Merchants are not just choosing whether to accept a new channel. They are choosing which economics they want to protect.
Savings versus monetization
Consumer agents can help users avoid waste and optimize spending. That is good for users. It can be bad for businesses whose economics rely on confusing terms, missed cancellations, or rewards structures that assume people will not optimize perfectly.
Practical Application
Trace every credential handoff
Before shipping any agentic payment feature, draw the exact flow on one page:
- where the credential is stored,
- who tokenizes it,
- what the merchant sees,
- what the issuing bank sees,
- who appears on the statement,
- and who handles refunds and disputes.
This exercise will usually reveal whether you are preserving the normal commerce lifecycle or only making the first payment succeed.
Scope delegated spend tightly
Do not start with blanket spending authority.
Ask for structured permission such as:
- product category,
- merchant or merchant type,
- spending cap,
- time window,
- and whether the permission is one-time or reusable.
The hotel example in the episode works because the agent is not acting without context. It has a budget and a clear task.
Decide merchant posture early
If you build for merchants, explicitly decide whether AI agents are a growth channel or a threat to a monetized interface.
That decision should affect:
- API strategy,
- checkout design,
- bot detection,
- partnership priorities,
- and whether you optimize for agent-readable flows or for deliberate friction.
A vague middle ground will usually produce the worst product, because both humans and agents get a compromised experience.
Preserve merchant follow-ons
Make post-purchase flows part of the original product spec.
Check whether a merchant can still do these things after an AI agent checks out:
- issue a refund,
- convert a purchase into a subscription,
- enroll the customer in loyalty,
- support a return,
- or re-engage the buyer later.
If the merchant only sees an intermediary's temporary credential, many of these flows will quietly fail.
Audit revenue tied to inattention
Review where your business depends on customer friction or forgetfulness.
Examples include:
- forgotten renewals,
- hard-to-cancel subscriptions,
- underused rewards,
- confusing fees,
- or poor price comparison.
If revenue depends on any of those, an always-on consumer agent is a direct strategic threat. That does not mean the business is doomed. It means pricing and retention will need to rest on real value instead of breakage.
Questions to Consider
- If an AI agent bought through your product today, could both the consumer and the merchant clearly explain who is responsible for a wrong item, wrong date, or duplicate charge?
- Does your business benefit more from human attention inside a monetized interface, such as ads or merchandising, or from being the easiest place for an AI agent to complete a purchase?
- Which parts of your checkout or subscription revenue rely on customer friction or inattention, and what happens when an always-on agent starts optimizing those away?
- If your product stores payment credentials, can those credentials safely serve multiple processors, agents, and merchant partners without locking the company into one provider?
Bottom Line
Agentic commerce is not mainly a UX problem. It is a trust, incentives, and infrastructure problem.
The products that win will be the ones that let agents spend while keeping the existing payment system's protections, merchant relationships, and liability boundaries legible to every party involved.
Case Studies Mentioned
Stripe Link at Walmart
Terry used an agent to buy outdoor lights from Walmart, but Stripe Link did not have a direct merchant relationship with Walmart. Stripe solved the immediate problem by using its own one-time virtual card to buy the product and then charging Terry separately. The transaction worked, but Walmart could not later convert the purchase into a Walmart Plus relationship because it did not have Terry's actual card. The lesson is that successful checkout is not the same as preserving the full commerce relationship.
Amazon blocks Muse
Alex points to Amazon blocking Muse as an example of why merchant behavior will vary. Amazon has a strong incentive to keep consumers inside a surface where browsing and ads can be monetized. An agent that skips straight to purchase may raise conversion but weaken a much richer economics stack. The lesson is that resistance to agents can be rational, not merely defensive.
Plaid and Chase precedent
Terry compares the coming merchant-agent negotiations to early open banking fights between fintechs and banks such as Chase. Over time, access was allowed, but often through controlled APIs, tokenization, and commercial terms rather than totally open scraping. The lesson is that agentic commerce will likely evolve through negotiated standards and constrained interfaces, not pure openness.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| VGS | Secures and vaults payment credentials and network tokens | Terry explains that VGS sits in the middle of merchants, wallets, agents, and networks, which gives it a clear view into how agentic payments are being built. |
| Stripe Link | Consumer wallet and checkout product from Stripe | It is a leading wallet in the wallet wars and the episode uses it to illustrate both the convenience and the limits of a walled-garden agent payment flow. |
| Shop Pay | Consumer wallet and accelerated checkout from Shopify | It is cited as another wallet with meaningful consumer and merchant scale that could become a default payment layer for agents. |
| PayPal | Consumer wallet and merchant payment network | It is used as a benchmark for installed account base and merchant reach in the race to become the wallet layer for agents. |
| Muse | AI shopping agent | It is discussed as an example of a new agent that becomes much more useful once it can access payment credentials, and as a product Amazon chose to block. |
| Instinct | AI shopping agent | Terry uses Instinct in his examples to show an agent completing tasks like product discovery and checkout on his behalf. |
People to Follow
Terry Angelos
VGS executive board member and payments operator with direct visibility into how wallets, merchants, and AI agents are handling credentials. He is especially worth following if you care about agentic commerce infrastructure, tokenization, and how legacy payment rails adapt to AI.
Alex Rampell
a16z general partner and co-founder of TrialPay, with a long history in payments and marketplaces. He is useful to follow for thinking on consumer payments, platform economics, and how infrastructure shifts change product strategy.
Notable Quotes
"agents being what I would call spectators to agents being actors, because they effectively have arms and legs, which are payment credentials." (Terry Angelos)
"we sort of have kicked off what I kind of call the wallet wars" (Terry Angelos)
"vaults are going from single player to multiplayer and that makes them a lot more valuable." (Terry Angelos)
"anything with a sort of breakage model in it is going to start to fall apart" (Terry Angelos)
