What Stripe Actually Shipped
On May 4, 2026, Stripe quietly launched a version of their Link wallet specifically designed for AI agents to hold and use. The capabilities, in plain language:
- **Agent gets a wallet you control.** Funded from your accounts, held in the Link product, accessible to a specific agent you authorize. - **One-time-use cards per purchase.** Each transaction gets a fresh card number scoped to that purchase. No card-on-file risk. - **Real-time approvals.** You get a notification when the agent is about to spend. You approve in the Link app. Spend happens. - **Machine payment protocols.** The wallet supports emerging machine-readable payment standards, not just human-style card transactions. - **Purchase history.** Every agent transaction is logged and auditable. - **Coming soon (per Stripe):** granular per-agent spending controls, digital currencies, and saved preferences for unattended purchases within rules.
Setup is the same .md spec file pattern that is becoming the default for the agent era: the agent reads link.com/skill.md and self-configures, or you run `npm install -g @stripe/link-cli && link-cli onboard` and walk through it interactively.
This is a real launch from one of the most credible infrastructure companies in the world. Agentic commerce stops being theoretical the moment Stripe ships the rails.
Why the Wallet Was the Missing Piece
Before this, every Sentie agent (and every serious custom AI agent from any vendor) had the same hard stop on the workflow: payment required. The agent could prepare the order, fill the cart, draft the invoice, identify the supplier. Then the workflow handed off to a human to check out, because card-on-file gives the agent too much spending authority and ad-hoc card entry requires a human at a screen.
This was the right hard stop. Agents in 2025 had no good way to spend money safely. The infrastructure for per-transaction human approval simply did not exist.
Stripe Link wallet for agents fills that gap. The agent now has a way to spend with human approval retained per transaction, not just per onboarding. That changes which workflows are viable.
What This Unlocks for Real Businesses
Five categories of custom AI agent become meaningfully more viable now that the payment substrate exists:
**1. Marketing-spend agents.** An agent that watches campaign performance across Google Ads, Meta Ads, and LinkedIn Ads and reallocates budget in real time within human-approved rules. Without a wallet, the agent could recommend and the operator had to execute. With a wallet, the agent shifts the budget directly while you get notified of every reallocation above a threshold.
**2. Procurement and replenishment agents.** Office supplies, software subscriptions, vendor renewals, consumables for a service business. The agent watches stock levels, identifies the right supplier, places the order. For HVAC and other field-service businesses, the most expensive version of this is parts ordering for trucks - dispatch agent identifies the part the tech needs, agent orders directly from the parts distributor.
**3. Inventory replenishment for ecommerce.** Agent watches per-SKU stock, demand patterns, supplier lead times, and reorders within rules. Most ecommerce stores have someone whose job is essentially this; with a wallet, the agent does it.
**4. Travel and logistics.** Within travel policy, the agent books flights, hotels, and ground transport for the team. Per-traveler limits set by the operator. The agent handles the booking-and-rebooking cycle that consumes hours per traveler.
**5. Subscription and vendor management.** Agent watches your subscriptions, flags renewals coming up, negotiates or cancels per your rules, pays renewals you approve. The category that AI most clearly improves over a human's attention span.
None of these were impossible before May 2026. All of them required a human checkout step that fragmented the workflow. The wallet removes the fragmentation.
What Stays Hard (and What You Should Watch For)
Three real risks that the wallet does not solve automatically:
**1. Agent fraud is a new category.** Adversaries are already working on prompts that trick agents into approving purchases they should not. The wallet limits the blast radius (one-time-use cards, per-transaction approval) but the operator still needs to actually read the approval prompts. Mindlessly approving "agent wants to buy X for Y dollars" 50 times a day creates the exact conditions for an attacker to slip a bad request through.
**2. Anomaly detection still needs you (or Sentie).** The wallet logs every transaction. Whether those transactions look right is your problem. A serious custom AI agent layer should sit on top of the wallet logs and flag spending patterns that look off - sudden spikes, unusual vendors, off-hours activity. Sentie's substrate includes this kind of guard rail; not every agent vendor does.
**3. Reversibility differs by transaction type.** Refundable purchases (most card transactions) can be reversed if the agent gets it wrong. Some transactions cannot. Set higher approval thresholds for irreversible spend categories and make sure your agent's playbook knows the difference.
These are real problems, but they are the kind of problems we know how to solve. They are not reasons to wait. They are reasons to deploy with a vendor that takes them seriously.
How Sentie Is Building Around the Wallet
We are adding Stripe Link wallet integration as a substrate-level capability, not just as another integration in our 248+ tool catalog. The difference matters: integrations let agents use a tool; substrate-level capabilities are part of the agent's default operating model.
What that means concretely:
- **Per-tenant Stripe Link wallet** as part of your [Business Brain](/blog/what-is-a-business-brain). Configured during the assessment phase, shared across all your specialized agents (sales, support, ops), with per-agent spending limits you set. - **Spending policy in the constitution file.** The plain-text spec your agent reads on every action includes the rules - max single transaction, allowed vendor categories, off-hours behavior, approval thresholds. - **Anomaly detection guard rail** layered on top of the wallet logs. The substrate watches transaction patterns and escalates anomalies to you before they post. - **Reversibility-aware playbooks.** Skills involving irreversible spend (wire transfers, certain B2B card transactions, anything where the recipient is hard to reverse) get higher approval thresholds by default and never run unattended even within rule.
We are dogfooding this on Sentie's own operations first. Our own marketing-spend agent (running on Sentie internally) will be the first production deployment of the wallet substrate. Once we have run it on ourselves for a quarter, we will document the results and open the capability to customers.
What Buyers Should Do Right Now
Whether you are a current Sentie customer or you are evaluating a custom AI agent vendor, three actions worth taking this quarter:
**1. Identify the workflows where payment is the bottleneck.** Walk through your team's actual day. Where does an otherwise-AI-shaped workflow currently stop because of a payment step? Marketing spend? Parts ordering? Vendor renewals? Travel? That list is your roadmap for agentic commerce in your business.
**2. Pressure-test your vendor on agent payment security.** Ask: how do they integrate with the Stripe Link wallet (or equivalent)? How are spending rules expressed? What anomaly detection is in place? Vendors that have not thought about this will tell you with their answer.
**3. Start small with a single agent payment workflow.** Pick one bounded workflow (e.g., software subscription renewals under $200 per transaction). Run it for a quarter with human approval on every transaction. Track what worked and what surprised you. Expand from there.
The pattern with agentic commerce is the same as with the rest of AI: the businesses that integrate this layer thoughtfully in 2026 will compound an operational advantage that catch-up later cannot replicate.