Link… allows you to connect various payment methods, track your spending, and view your recurring subscriptions. It also lets you integrate your AI agents so they can spend on your behalf, securely.
For the buyer, this is the first agentic-commerce story that actually feels safe. Same wallet, same approval UX, regardless of which agent or which shop. You’re not learning a new trust model per vendor. You’re learning one, Stripe’s, and reusing it everywhere. That’s a big deal. Every previous “let the AI buy things for you” pitch asked the user to either hand over a card or trust some startup’s bespoke flow. Now there’s a familiar shape.
users will first grant their agent access to the Link wallet via an OAuth (standard authentication) flow. The agent can then create a spend request, provide you with the context, and wait for your approval.
Approve, deny, see the context. That’s the entire mental model. A grandmother could learn it. That’s what makes it real.
Instead of giving an agent access to your payment credentials, users can either give agents programmatic access to Link, which provides a one-time-use card, or they can use a Shared Payment Token (SPT), which is backed by payment cards and banks.
For the seller, the integration story is suddenly trivial. If you already take Stripe, you already take agents. One implementation, one point of contact, one risk profile. No bespoke AI-checkout protocol per agent vendor, no fragmented standards war. Plumbing solved.
And here’s where most coverage stops. Plumbing is the easy part.
Developers and businesses building agents or AI personal assistants can also use Link’s wallet instead of building their own wallet from scratch.
Stripe quietly handed sellers a much bigger bill: your shop now has to be legible to a non-human buyer. Three new workstreams nobody’s budgeting for yet.
One: your catalogue has to be LLM-optimised. Not SEO. SEO is for humans skimming. Agents read structured data, schema.org, clean product feeds, honest specifications. Vague descriptions, hero-shot-only product pages, and “ask in store for details” all become invisible. If the agent can’t parse it, it can’t buy it. Which means it doesn’t buy it from you.
Two: your logistics has to be agent-grade. A human buyer forgives a vague delivery date. An agent comparing three sellers in real time does not. Stock accuracy, delivery windows, return policy: all of it needs to be queryable, machine-readable, and honest. A shop that lies about “in stock” loses to one that doesn’t, every time.
Three: your customer care has to handle conversations the human never saw. Package didn’t arrive. Size is wrong. Agent ordered the blue one and the customer wanted navy. Who handles that? Ideally another agent on the seller’s side, because if a human has to stitch together what the buyer’s agent did and what the buyer actually wanted, support cost eats the margin. Most sellers will discover this the expensive way.
in the future, Stripe says it will expand its controls so users can set their own spending limits, or even choose when their agents can act without approval.
When approval-free purchases become normal, the support load shifts from “did the human mean this?” to “did the agent get this right?” Different question. Different tooling. Refund flows, dispute flows, return flows: all designed for human regret. Now they have to handle agent error too. A different shape of mistake.
The boring part is the interesting part. Stripe just removed the friction from agentic commerce. What’s left is everything around it: catalogue work, logistics work, support work. Sellers who treat that as the new baseline will quietly own the next decade of e-commerce. Anyone who treats Link as “we just turned it on” will wonder why the agents keep buying somewhere else.
So: who’s helping the SME shop owner do that work? Stripe just made the easy part easy, and the hard part suddenly visible.
🌈
techcrunch.com