Link is interesting to me for a simple reason: agents will need to charge.
Not just recommend things. Not just fill out forms. Charge. Pay. Finish a purchase.
That is where it gets weird.
An agent can help you pick a flight. But when it is time to pay, you do not want to hand it your card. You also do not want to type it every time. That is friction. And it is a bad boundary.
Link can be a clean way to solve that part.
What Link is
Link is Stripe's wallet for faster online checkout.
It stores payment methods, billing addresses, and shipping details. It can work across businesses, browsers, and agents.
Its relationship with Stripe is direct: Link runs on Stripe's payments infrastructure. In a Stripe Checkout integration, you can offer Link as a payment method so the user can pay with information they already saved.
The useful part for agents is simple: the agent prepares the purchase, but the user approves the payment. The credentials are not exposed.
More context:
The problem is not checkout
The problem is trust.
An agent that answers questions is easy to understand. An agent that buys things for you is a different category.
Buying, booking, renewing, subscribing. All of that has consequences. There is money involved. There are mistakes. There are refunds. There is support.
That is why the design matters.
The agent should be able to prepare the purchase. But it should not keep the payment details.
Link already stores payment methods, addresses, and checkout details. It also makes approvals fast. If that connects well with agents, the flow gets much more reasonable:
the agent does the work, the user approves the charge, and the card stays out of the agent's reach.
What it looks like
A basic flow:
- The user asks the agent to buy or subscribe to something.
- The agent prepares the order.
- Your system creates the charge with Stripe and offers Link.
- The user reviews and approves.
- Link processes the payment.
- The agent continues with the confirmation.
That is the part I like.
The decision can be assisted by an agent. The authorization does not have to be.
The agent can research, compare, and prepare. The user still says yes.
Where it makes sense
I do not think this is useful for every chatbot. It is useful when the agent is close to a real transaction:
- A shopping assistant.
- A support agent that charges for an upgrade.
- A B2B flow that pays approved vendors.
- A marketplace where the agent coordinates order and payment.
- Personal automations that book or renew something.
The important part is not that the agent spends on its own.
That actually sounds pretty bad.
The important part is that it gets to the point of payment and asks for approval.
Why it matters
Agents will become more useful when they can finish complete tasks.
But many tasks end in a payment.
That is where Link fits well. Not as some grand futuristic idea. More like boring infrastructure, which is exactly what you want when money is involved.
Less demo. More done.