Stripe’s Link agent wallet added recourse. The seller is still the agent’s word

Published October 2, 2026 · Last reviewed October 16, 2026

“openclaw-jane wants to spend $180.” Under that line, the sample approval screen on Link’s page for agents shows a hotel called Triplo, the domain triplo.com, a queen room in New Orleans and two buttons.¹ When the agent pays with a one-time card, the default path, it writes most of that screen itself. Stripe’s documentation has the agent pass the merchant’s name and URL, plus at least 100 characters on what it is buying and why. The consumer sees all of it before approving.²

Stripe’s post of September 29, 2026 improves what happens around that tap. An agent can ask for a higher amount after approval, get guidance past a stalled checkout, and read the consumer’s spending history with permission. Eligible purchases carry free purchase protection, with Meta’s Muse the first agent to offer it. Stripe says agentic purchases made with Link rose 38x over the past month.³

Read against Stripe’s documentation and Link’s terms, two of those items are narrower than they sound. The consumer still approves every increase, and the protection attaches to the goods, not to the agent’s choices. Underneath both is a third gap: on the card path, the seller is whoever the agent says it is.

What Stripe shipped on September 29

Stripe launched Link’s wallet for agents at its Sessions conference on April 29, 2026.⁴ Each purchase starts as a spend request. Only after the consumer approves does the agent get a credential, a one-time-use card or a token for supported Stripe checkouts, never the real card details. Agent payments are open to consumers in the US and Canada.² The September 29 post, by Link product manager Dan Hill, adds three things.³

Incremental authorization and checkout guidance. When the final price comes in above the approved amount, the agent can ask to raise it before using the card. Stripe’s example is a checked bag added after the fare. When a payment stalls, Link returns a next action. After a 3D Secure challenge the same request resumes. A declined card, or a charge attempted above the approved amount, means a new request.²

Financial insights. With permission, an agent can read transactions, balances and account details from the accounts a consumer shares, and the consumer can revoke access in the Link app. It runs on Financial Connections, which Stripe says reaches over 97% of US bank accounts, and is US-only.⁵ For now, Link’s agents page lists it only for agents built on Link’s command-line tool.¹

Purchase protection. Eligible Link purchases carry free cover for eligible US residents: up to $500 for damage, theft or loss, up to $500 after a price drop, up to $250 in return fees, and up to $1,000 when a business refuses a return.⁶ For Muse this is not new: Link’s protection terms took effect on September 8, the day Muse launched, with a section for Muse purchases.⁷⁸ The September post presents it as a wallet feature more agents will get.³

Stripe gives Muse, Grok Bot and Instinct as examples of popular agents on the wallet.³ Grok Bot comes from SpaceXAI, the name xAI has carried since July.⁹ Browser Use connected on September 3.¹⁰

Raising the amount still takes a tap

Incremental authorization is the most practical item in the release, and the easiest to misread.

Before it, an agent that found a higher total had to start over with a new request for the correct amount. Stripe says the feature saves restarting the purchase and placing a second hold on the consumer’s card.²³ The agent now updates the existing request and asks again. The documentation is direct: “Raising the amount requires the customer to approve again.” Increases are not guaranteed, and the documented fallback is to cancel and start over.²

So the tap stays. What goes away is the restart. That matters, but it is not the scenario people picture, an agent finishing a booking while its owner sleeps. A fare that rises at 2 a.m. still needs someone awake to approve it before the request lapses.

That job belongs to the controls Stripe lists as next. The consumer would set a budget, and the agent could finish a task within it without approving each transaction. Stripe’s example is a pair of limited-release shoes going on sale at 1 a.m.³ The April launch post already said Stripe planned to let people set spending limits and choose when agents can act without approval.⁴ Link’s agents page still lists those controls as coming soon.¹

The protection covers the goods, not the agent’s choices

Stripe frames purchase protection as recourse “if an agent makes a mistake.”³ The cases it lists are close to Link’s terms. The framing is broader.

Link’s terms describe insurance programs with Stripe as the named insured, say Stripe is not an insurer, and make eligibility depend on the merchant.⁷ The coverage guide is more specific: damage and theft sit under a policy from an AIG company, and Stripe itself provides the other benefits. Either way, the cover attaches to the item: broken, stolen, cheaper a week later, costly to return, or refused for return. Even there the fine print is tighter than the marketing. Stripe’s post lists lost items, but the guide excludes items that are “lost or mysteriously disappear,” and a theft claim needs a police report.¹¹

The agent’s judgment is a different matter. Link’s help page makes the consumer responsible for the agent’s transactions, including a different product, merchant or option than intended: “If the agent made an error, contact your agent platform.”¹² The terms go further. The consumer answers for agent-initiated transactions “as if you had taken the action yourself,” including ones caused by bugs, hallucinations or misinterpretations, unless Stripe’s technical failure caused them. Those transactions also fall outside Link’s protection against unauthorized transactions.¹³ A wrong item can still go back with return fees covered, if the merchant takes returns. A wrong seller stays with the consumer.

American Express draws the line elsewhere. Its Agent Purchase Protection, announced in April, says Amex will stand behind the consumer when an agent makes a mistake, provided developers have registered that agent with Amex.¹⁴ Amex ties recourse for agent errors to knowing who the agent is.

Who the seller is, and who says so

Link answers the seller question differently depending on the credential. Some Stripe-hosted checkouts accept a Link Pay Token, and some sellers accept Stripe’s shared payment tokens. On both, Link resolves the seller server-side and shows the consumer the canonical merchant before approval.²¹⁵ Everywhere else, the agent asks for a one-time virtual card. Stripe’s Muse announcement draws a similar line: Link checkout at more than a million businesses that accept it, and a single-use virtual card for the rest.¹⁶

On the card path, the agent supplies the merchant’s name and URL, and nothing in Stripe’s documentation says Link checks them. The card is limited to one use, a set amount and 12 hours, and it “works anywhere,” including at sellers that do not use Stripe.¹⁵ The main check on the seller is the consumer, reading an agent-typed name in a notification.

After the tap, Stripe sits in the middle. Link’s terms say an agent purchase can appear on the consumer’s statement as LNKAGNT* plus the merchant name. In some cases Stripe pays the merchant with a separate payment method, the consumer reimburses Stripe, and the underlying card’s benefits do not apply.¹³ In that arrangement, the seller is paid by a card Stripe issues and the consumer’s bank sees Link.⁴ Stripe is the party with a view of both sides.

Card networks work the other direction. Visa’s Trusted Agent Protocol, introduced in October 2025 with Stripe among the early partners it credited, lets approved agents pass signed information so merchants can tell them from malicious bots.¹⁷ Stripe’s April post listed support for agentic tokens, the card networks’ credentials for agent purchases, as coming soon for Link.⁴ Such schemes tell a merchant about an agent. They do not tell the agent who the seller is, a gap that runs through the rest of the agent payments stack.

Muse shows both halves. It was the first agent covered by Link’s protection, and on September 22 PayPal announced a partnership that adds its own checkout to Muse, outside Link’s cover.¹⁸ In the same month, Amazon blocked Muse over identifying the agent at the door. A payment credential tells neither the seller who the agent is nor the agent who the seller is.

The planned budgets sharpen this. Today the approval screen is where a human reads the seller’s name. Under a budget, an agent could pick the seller at night and nobody would read the screen. Some people want that mode: in research released September 30 by the lender Synchrony and Oxford Economics, 43% of 2,000 US consumers surveyed in May were comfortable letting AI buy up to a preset limit.¹⁹ Caps bound your own agent, not the party it deals with, as with Circle’s delegation controls. The machine-side pieces, ERC-8004 for on-chain reputation, Web Bot Auth for signed requests and A2A Agent Cards for discovery, answer a machine’s questions. The person approving a $180 hotel still reads a name the agent typed.

What 38x does and does not measure

Stripe’s 38x is a ratio, reported by the vendor, with no base. Stripe gives no transaction count and no dollar volume for agentic purchases through Link. Link has over 300 million users by Stripe’s count, but that is the wallet, not its agent traffic.¹⁶ The month also included Muse’s launch, so one new agent could account for much of a jump from a small base. Stripe does not break the figure down.

One signal outside Stripe points the same way without settling the size. Instinct’s founder told an interviewer the service is approaching a billion dollars in annual transactions, not necessarily through Link, and TechCrunch noted he did not explain the calculation.²⁰ Treat 38x the way this blog treats tracker-attributed x402 volume figures: agent checkout is moving toward default, and the market’s size is unknown.

How this connects to Tobira

Tobira’s contribution sits on the human side of that problem, next to the payment rails rather than in place of them. An @handle gives the person or company behind an agent an address a human can recognize and another agent can look up. A Site Agent makes a company website agent-addressable and networked, so a visiting agent can ask it questions and reach the owner. Mutual reveal records that both humans agreed before contact details change hands, argued in why agent networks need mutual reveal. Tobira issues no cards, holds no balances and handles no claims. Claiming a @handle on Tobira is free, and the network had 729 agents per Tobira internal data, September 2026. Tobira Site Agent has a free plan on your own AI key and paid plans from $49 a month (as listed on tobira.ai/agentic-site/ on 16 October 2026). Next time an agent asks to spend $180 at a hotel you have never heard of, what on the screen tells you it exists?

What to remember

• On September 29, 2026 Stripe described incremental authorization, checkout guidance, financial insights and purchase protection for Link’s wallet for agents, and 38x monthly growth in agentic purchases.³
• Raising an approved amount still needs a second approval. It saves a restarted purchase, not the tap.²
• The protection covers eligible goods for US residents, partly through insurance. Link’s terms leave the agent’s choice of product or merchant with the consumer.⁷¹²
• On Link Pay Token and shared-payment-token checkouts, Link resolves the seller. On the default card path, the agent supplies the merchant’s name and URL, and the card works anywhere.²¹⁵
• Planned budgets would let agents buy without approving each purchase. They would also remove the point where a human reads the seller’s name.³

Frequently asked questions

What did Stripe add to Link’s wallet for agents on September 29, 2026?

Incremental authorization, so an agent can ask to raise an approved amount, and next-step guidance for stalled checkouts. The post also covers permissioned financial insights through Financial Connections and free purchase protection on eligible purchases, with Meta’s Muse first. Stripe reports that agentic purchases made with Link rose 38x over the past month.³

Can an agent on Link spend more than I approved?

No. To raise an approved amount, the agent updates the spend request and you approve again. Budgets that would let an agent spend without approval for each purchase are planned, not available.²³

Does Link’s purchase protection cover an agent’s mistakes?

It covers eligible goods: damage, theft, price drops, return fees and refused returns, for eligible US residents. It does not cover the agent’s choice itself. Link’s help page makes you responsible for the agent’s transactions, including a different product, merchant or option than intended, though a wrong item may still be returnable with fees covered.⁶¹²

Which AI agents use Link’s wallet for agents?

Stripe names Meta’s Muse, Grok Bot from SpaceXAI (formerly xAI) and Instinct, and Browser Use has also integrated it. Other agents can connect through Link’s command-line tool, SDKs or a local MCP server. Agent payments are available to consumers in the US and Canada.³¹⁰¹⁵²

How do I know the seller my agent picked is real?

On checkouts that take a Link Pay Token or a shared payment token, Link resolves the seller and shows you the canonical merchant before approval. On the default one-time-card path, the merchant name and URL come from the agent, and Stripe’s documentation does not say Link verifies them. Check the domain, and treat an unfamiliar seller as unverified.²¹⁵

Sources

Figures, dates and feature statuses were checked against the sources above on October 2, 2026. Where Stripe describes a control as planned or coming soon, this article does not treat it as shipped.