Stripe on April 29 launched a digital wallet purpose-built for the era of autonomous AI agents, letting consumers grant a software agent permission to shop, book reservations, and pay bills without ever exposing the underlying card or bank credentials.
The product, called Link’s wallet for agents, is the headline consumer-facing piece of an unusually large product release: 288 new features and APIs unveiled at the company’s annual Stripe Sessions conference in San Francisco. CEO Patrick Collison framed the bundle as the “economic infrastructure for AI,” arguing that “in the not-too-distant future agents will account for most transactions online.”
The launch turns Link — a Stripe-owned consumer wallet with more than 250 million users globally — into the first mainstream payment surface explicitly designed for the agent-to-agent era of e-commerce that Stripe and its rivals have spent the last 18 months trying to build out.
Source: Stripe
How the Agent Wallet Actually Works
Stripe’s pitch is that consumers should not have to choose between agent convenience and credential security. In Stripe’s own description of the architecture , users first grant an agent — for example, the OpenClaw shopping assistant — access to their Link wallet through a standard OAuth flow. The agent can then create a spend request scoped with merchant, amount, currency, and a natural-language description of what it’s about to buy.
The user receives a notification on web, iOS, or Android to review and approve the transaction. Only then does Link issue either a one-time-use virtual card or a Shared Payment Token (SPT) — a scoped credential backed by the cards or bank accounts already in the wallet — that the agent uses to complete the purchase. The agent never sees the underlying payment method.
Under the hood, the consumer wallet sits on top of a developer product Stripe is calling Issuing for agents, a set of primitives for single-use cards, transaction-level spending controls, real-time authorization, and full audit trails. Businesses building their own agent-facing assistants can use it to skip building wallet infrastructure from scratch.
For now, Link’s agent wallet supports cards and SPTs. Stripe says support for “agentic tokens, stablecoins, and other types of payments is coming soon,” and the company has previewed expanded controls that will eventually let users set spending limits or pre-authorize agents to act without per-transaction approval.
A Coordinated Push Across the Payments Stack
Link’s agent wallet does not stand alone. The same Sessions 2026 announcements included a partnership with Google to let businesses sell inside Google’s AI Mode and the Gemini app via the Universal Commerce Protocol, the Machine Payments Protocol (co-authored by Stripe and the Tempo blockchain) for direct agent-to-business microtransactions, and an Agentic Commerce Suite that funnels Wix, BigCommerce, and WooCommerce merchants into AI shopping surfaces with a single integration.
Will Gaybrick, Stripe’s president of product and business, framed the company’s posture in characteristically blunt terms: “If AI can solve Nobel level physics problems but can’t buy a domain, something’s gone wrong. Our mantra: empower agents.”
That mantra has competitive teeth. Stripe already has partnership announcements with OpenAI, Microsoft, and Meta , and the Visa Intelligent Commerce and Mastercard Agent Pay programs both ride on Stripe’s SPT primitive. Etsy and URBN — parent of Anthropologie, Free People, and Urban Outfitters — are listed as early adopters, alongside Affirm and Klarna for buy-now-pay-later flows.
The race to define the agentic-payments substrate is now genuinely crowded. Anthropic’s own Project Deal effort to let Claude agents trade real goods launched earlier this month; Tempo, the Stripe- and Paradigm-backed payments blockchain underpinning the Machine Payments Protocol, went live in March; Mastercard and Stripe announced a separate integration with the Wizard agentic-shopping platform on the same day. Whoever owns the wallet, token format, and merchant catalog at the bottom of the stack will collect a small percentage of every agent-led purchase for the next decade.
What Could Still Go Wrong
The friction Stripe is engineering away — manually entering card numbers across dozens of agent surfaces — is also the friction that has historically capped consumer fraud. A wallet that can hand out one-time-use cards programmatically, on instructions from software acting on a user’s behalf, raises the question of who is accountable when AI starts transacting and a charge later turns out to have been induced by a prompt injection, a hijacked agent, or simple model error.
Recent incidents have shown that agentic AI failures can cascade in seconds when an autonomous system has live credentials. Stripe’s design — per-transaction approval today, scoped tokens with merchant and amount limits, OAuth-style revocation — is more conservative than what industry roadmaps describe further out. By stage five of Stripe’s own published agentic-commerce framework, agents act fully autonomously based on user history. The gap between today’s approve-everything model and that endpoint is where most of the consumer-protection, dispute-resolution, and agent oversight hard problems live.
What to watch next: how quickly Stripe’s “soon” stablecoin and BNPL token support actually ships through Privy and Tempo; which agent platforms beyond OpenClaw integrate Link first; whether Apple Pay and Google Pay respond with their own agent-mode credentials; and whether Visa and Mastercard’s network-issued agentic tokens become the default rail or remain one option among many. The plumbing for an internet that’s increasingly built for AI is being poured in real time, and at Sessions 2026, Stripe made the strongest claim yet that the wallet at the center of it will be theirs.