"It's easy to underestimate the implications of a near future where agents are autonomous economic actors, responsible for most internet transactions. And just as agents work faster than we do, they're going to spend money a lot faster too."

That is Will Gaybrick, Stripe's president of technology and business, on stage at Stripe Sessions on April 29, 2026. The line set up the Link announcement, but the framing that stuck came from a guest. Gaybrick asked Ginger Baker, who runs fintech and payments product at Meta, what sounds radical today and will look obvious in two years. Her answer, from Stripe's own transcript: payments will "pivot from being a moment to being a policy."

That is the sentence to keep. The Link expansion is not really about a wallet. It is about what the wallet enforces.

The credential was the old product. The policy is the new one. Stripe told the rest of the industry that the rules engine sitting between a user and an agent is where the next decade of consumer payments lives.

The mechanic is straightforward. A user grants an agent access to Link through an OAuth flow. When the agent needs to spend, it creates a spend request with the amount, the seller and a plain-language description of the purchase. The user gets a notification, approves, and Link releases a one-time-use card or a shared payment token scoped to that purchase. The agent never holds the user's real card or bank details, Stripe wrote at launch, and TechCrunch confirmed the flow.

Link itself is a consumer wallet that Stripe said had more than 200 million consumers at launch. It stores cards, bank accounts and subscriptions and now tracks what connected agents spend across web, iOS and Android. The agent integration sits on top.

Stripe did not launch with a roster of big-name assistants. Its blog named one example agent, OpenClaw, and the keynote demo used OpenAI's Codex as the buying agent. The launch was a wallet opening a door, not a partnership announcement.

Consumer-Side, Not Merchant-Side

This is the announcement to separate from Stripe's Google partnership announced the same day. The Google deal extends Stripe's agentic checkout to merchants on the AI platform side, inside Gemini and AI Mode. Link extends authorization to consumers on the user side. Different problem, different surface.

Merchant-side agentic commerce asks: how does an agent pay a checkout that was built for humans? Consumer-side asks: who lets the agent spend in the first place, and on what terms? Both layers have to exist. Stripe is building both, which is why the company shows up at every angle of the agentic stack.

Treating the two as one announcement misses what is new. The Link integration is the first time a consumer wallet has shipped an agent authorization model that is more than a card-on-file shortcut. The user is in the loop on every transaction, by default.

The Shift from Credential to Policy

Per-transaction approval is the launch product. It is not the destination.

Stripe's launch post described the next steps: user-defined spending limits, agents acting under specified conditions without an approval tap, and visibility on what each agent is doing with the credential. That is the policy engine. At launch the user clicked yes on every spend. The plan is for the user to write a rule and the wallet to enforce it.

This is the part Baker named. "Payments will pivot from being a moment to being a policy" describes a shift the industry had been edging toward without a clean phrase. The moment is the checkout button, the tap, the approval. The policy is the standing instruction the wallet runs against every agent request: limits, categories, time windows, merchant allow-lists, counterparty rules. Her example on stage was an agent told to spend only up to $50 on groceries this week.

The wallet that ships the better policy engine will end up holding the consumer agent relationship. Not the wallet with more cards on file. Not the wallet with the slickest UX. The wallet that makes it easiest to express "yes, but only under these conditions" and trust that the conditions will be honored.

What Was Missing at Launch

The policy engine was the roadmap, not the launch. That was a substantial gap, and the three parts of it are the same three gaps I keep finding across the delegation problem.

There was no published rule grammar. Limits, categories, time windows and agent-specific scopes were listed as future capabilities without specifications. A user who wanted to set Baker's grocery rule could not. The wallet did not know how to express it. The agent did not know how to read it.

There was no agent identity layer in Link's announcement. The OAuth grant authorizes a session, not a verifiable agent. Stripe is trusting the platform behind the agent to vouch for its own software. That is a workable shortcut for a handful of early agents. It is not a model that scales to a thousand third-party agents, which is the question who authorized the agent keeps asking.

And there was no clarity on revocation. If a user's agent is compromised, or behaves outside its scope, the path from "stop spending" to "stopped spending" was not described. In a per-transaction model the question matters less because each spend is gated. Once the wallet starts honoring policies without per-transaction approval, revocation latency becomes the security boundary.

Six months on, the policy engine is half built, and the half that exists is instructive.

Stripe's current developer documentation shows the first standing rule. An agent can request a recurring virtual card with a spending limit per day, week, month or year. The user approves the terms once, and the seller charges against the card inside that cap until the user cancels it. Canceling revokes the credential. That is a policy, approved as a moment. It is narrow, one merchant and one cap, but it is the first time Link has enforced a rule rather than a tap.

The approval window is 30 minutes. Each wallet carries per-transaction, daily and 30-day agent limits that the agent can read before it asks, and Link's consumer help pages say the limit depends on purchase behavior and the agent in use, resets daily and cannot be raised. Disconnecting an agent expires every open request and every card it generated. Those are the revocation path and the rule grammar in their first versions. Neither is a published specification an agent on another wallet could read.

The volume is small but real. Upstarts Media reported on October 18 that Link powers about 200 agents, up from roughly 70 when Stripe ran its first analysis in September, and named Meta's Muse and Grok Bot among them. The Decoder picked up the same report: September agent transactions were 38 times the earlier level and volume 20 times, from a base Stripe will not disclose. Dan Hill, who runs Link, said the volume is "becoming very material." Material to a wallet of around 300 million users still means a rounding error.

The agent identity gap has not moved. Access is still an OAuth scope granted to an application. Nothing in the flow proves which model, which operator or which version is holding the credential. The policy is getting richer. The party it binds is still a session.

Why This Is the More Important Stripe Announcement

The Google partnership got the headlines. It deserved them. It is also a defensive move, the kind of distribution play Stripe runs every couple of years.

Link's expansion is offensive. The company that owns the consumer policy layer for agentic spend owns a relationship that PayPal, Apple Pay and the bank apps have been competing for since smartphones. None of them had shipped an agent-authorized wallet that is consumer-default. Stripe did, and named the policy framing while doing it.

The race for the agent-addressable wallet is now on terms Stripe has set. That is worth more than any single platform integration.

What to Watch

Three signals will tell me whether Link's policy layer becomes the standard or the example.

The first is the rule grammar. The recurring card is a cap per merchant per period. The next step is a rule an agent can read before it acts: categories, allow-lists, conditions. Whatever Stripe ships as that specification will become the de facto interface every agent has to honor. A proprietary JSON shape ships fast. An open standard ships slow but ships everywhere.

The second is the bank response. JPMorgan, Wells Fargo and the European tier-one banks have been quiet on consumer agent authorization. Link's framing forces them to either build a comparable policy layer or cede the consumer-agent relationship.

The third is the first revocation incident. A user whose agent goes rogue and whose Link policy fails to stop the spend will define the SLA for the entire category. The 30-minute window and the disconnect button are the current answer. Nobody has tested them under a real attack in public yet.

The wallet that enforces the policy is the wallet that wins. Stripe told the rest of the industry that is the game now. The question is who else is willing to play it.

When your wallet enforces a policy your agent must obey, who writes the policy?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.