Stripe Sessions 2026 was the company's declaration that agents are no longer hypothetical. The keynote framed agents as autonomous economic actors responsible for the majority of internet transactions. Across 288 product announcements, this is the slice that matters for agentic commerce — the surfaces, the protocols, the wallets, the guardrails, and what Stripe expects developers to build next.
Five levels of agentic commerce
John Collison opened day two with a useful framing for the entire suite of announcements: agentic commerce sits on a spectrum of autonomy, from Level 1, where software fills checkout forms on your behalf, to Level 2, where a shopping assistant reasons within constraints (the ChatGPT and Wayfair-style query patterns), all the way to Levels 3 to 5, where agents discover products, decide for themselves, and complete purchases without human input. We previously broke down each level in detail in our explainer on the five levels of agentic commerce.
Stripe’s pitch is that the rest of the keynote is infrastructure for every level of that spectrum, so merchants do not have to predict which mode of agent shopping will dominate before they start integrating.
Buy inside the four AI surfaces
Stripe now powers product discovery in ChatGPT through the Agentic Commerce Protocol (ChatGPT discontinued its original Instant Checkout in March 2026 and pivoted to discovery plus merchant redirect), checkout in Microsoft Copilot across 500,000+ US merchants, and inside Meta’s native ad checkout on Facebook and Instagram. The headline reveal at Sessions was a new Stripe and Google partnership that lets customers buy directly in AI Mode in Search and the Gemini app over UCP, with Quince, Fanatics, and JD Sports launching as the first retail partners. With both ACP and UCP rails now covered, merchants no longer have to pick a protocol to bet on.
Agentic Commerce Suite
The Agentic Commerce Suite is a single integration that exposes your catalog to every AI surface Stripe partners with, while you remain merchant of record and Stripe handles checkout, payments, and fraud prevention end to end. Catalog and per-agent access can now be managed directly from the Stripe Dashboard.
At Sessions, Shopify joined as the preferred catalog provider for retailers using the suite, Wix and Commercetools came on as launch platforms, and Best Buy and Kate Spade were named as the first enterprises live in production.
Checkout Studio
Stripe also rebuilt how merchants ship checkout itself. Checkout Studio is a new place inside the dashboard for configuring, A/B testing, and optimizing checkouts side by side, with a visual builder, a customizable embedded form, and the ability to compare conversion across payment methods in one view.
The agent-relevant detail is the Copy for LLM button. Once you have configured a checkout in the visual builder, a single click hands a structured spec to your coding agent, and a companion prototyping tool can scan your GitHub repo, identify the existing Stripe integration, and open a pull request with the changes through GitHub Actions.
Order Intents and Machine Payments Protocol
The Order Intents API lets you hand an agent a URL and a budget, after which the agent reads the page, parses variants, checks discounts, shipping, and tariffs, and completes the purchase end to end. For services that want to sell directly to other agents, the Machine Payments Protocol, an open standard Stripe co-built with Tempo, lets any endpoint tell an agent over HTTP that payment is required and how to pay, with Visa now running an MPP validator alongside Stripe and Standard Chartered.
The Order Intents demo from Sessions 2025, where an agent buys products end to end without a human filling forms:
And the Machine Payments Protocol introduction from Sessions 2026, including the agent-to-agent autonomous payment that follows immediately after:
Link Wallet for Agents
Link Wallet for Agents lets agents spend on your behalf without ever seeing the underlying credentials, with the user approving each purchase from the Link app. Link now reaches more than 250 million consumers globally, and Stripe added UPI in India (500M+ users), Pix in Brazil (160M+), and stablecoins as supported methods, all sitting behind the same agent wallet.
The keynote demoed all of it live with a fully autonomous two-agent transaction on stage: Claude built and deployed an API review app through Stripe Projects, Codex independently discovered it, recognized that it accepted MPP, and paid the $2 fee through Link, completing the first end-to-end agent-to-agent value exchange Stripe has shown publicly.
Agents buying from agents
The clearest evidence that this is no longer a theoretical roadmap came in John Collison’s day two talk. He cited cumulative downloads of payment-related skills on OpenClaw at 125,000 in twelve weeks, then ran a live demo of Claude autonomously buying Alpha Vantage market data for four cents over the Tempo CLI, generating a full research report from the data, and then publishing the report itself for sale so other agents and humans could buy it.
Stripe also called out an early ecosystem of services already accepting agentic payments: Parallel and Browserbase for agentic web browsing, and Postal Form, where an agent can mail a physical letter on your behalf. The takeaway from the segment was straightforward: software buying from software is the corner of agentic commerce that is already moving fastest, and there is a real first-mover advantage in being one of the services agents can transact with.
Tempo and streaming stablecoins
Tempo is the blockchain Stripe incubated with Paradigm specifically for payments and agentic finance, and mainnet just went live. The customers Stripe disclosed at Sessions are DoorDash for global Dasher payouts, Kalshi for issuing its own KLA USD stablecoin, Visa for running MPP and operating a validator alongside Stripe and Standard Chartered, and Stripe itself for powering Treasury financial accounts.
In a live on-stage demo, Stripe showed agents streaming sub-cent stablecoin micropayments per token consumed, settled on Tempo in real time, with Metronome (acquired earlier this year) handling the metering, rating, and alerting. The result is an entire class of pricing models that fiat rails cannot support, including paying per token as it burns.
Radar for AI businesses
Stripe’s fraud product also got an overhaul aimed squarely at AI-native businesses, where the unit being stolen is no longer just money but tokens of compute. Radar now covers three fast-growing patterns: multi-account abuse (Stripe says one in six AI signups is involved), free trial abuse (which more than doubled in the last six months and burns real inference cost on every fake trial), and pay-as-you-go abuse where customers consume thousands of dollars of tokens and never pay the bill.
Radar coverage was extended across every Stripe payment method, made available off-Stripe through an API for unified scoring, and paired with a new custom fraud models capability where merchants can pass their own product, catalog, and loyalty signals and Stripe trains a model fitted to them, with early adopters seeing at least 15% more fraud caught at the same false-positive rate.
Stripe Console
Stripe Console is a new agentic expert that lives inside your Stripe account, currently in preview. In the on-stage demo, Console was asked “how is my business doing month to date,” and it pulled live data, charted revenue and conversion by country, flagged that Brazil was underperforming, recommended enabling Pix, and offered to set up an A/B test on confirmation, all in a single flow.
Stripe also opened all of its Treasury APIs over MCP, then demoed the full loop by paying a $1,300 PDF invoice end to end from inside ChatGPT through the Stripe MCP server.
Developer surface for agents
The developer keynote spent most of its runtime making Stripe itself more readable and operable by agents. The Stripe MCP server picked up an Integration Recommender Tool that picks the right products based on the merchant’s geography, Stripe Agent Skills now load Stripe expertise directly into the agent’s context window before it writes any code, and the Stripe CLI has been repositioned as the primary agent interface, with new commands that let an agent provision a sandbox and fetch and store API keys without breaking flow.
Stripe also released an Integration Benchmark that measures whether an agent can complete common Stripe tasks correctly, made all of its docs available as raw markdown by appending .md to any URL, and added a Copy for LLM button on Blueprints that hands an agent a structured set of integration steps. Stripe Projects, the vibe-deployment product that lets agents provision Vercel, Supabase, and Runloop and navigate dashboards on their own, also moved from private preview to general availability.
Stripe paired the launches with internal stats that make the direction unambiguous. Its in-house coding agent, Minions, ships more than 1,000 production pull requests every week, 91% of Stripe engineers now use AI tools daily, and 57,000 PRs were created with AI assistants at Stripe in the month of March alone (up 140% in three months). Externally, agent traffic to Stripe’s documentation went from under 5% to nearly 40% in a year, and Stripe expects agents to read more of its docs than humans within months.
Agent guardrails
Stripe also formalized agents as a first-class actor type inside merchant accounts. Agent-tagged API keys give every agent its own identity, observability, and permissions, while approval rules layer safe-by-default thresholds across the API, MCP, and dashboard so any sensitive action can require a human in the loop regardless of who initiated it.
In the demo, an agent attempted a $250 refund, hit the $100 threshold, received a structured error explaining the rule, called the approval-request endpoint on its own, and waited for a human in the dashboard to sign off before completing the action. Ginger Baker, VP Product at Meta, summed up the broader direction on stage: “Payments will pivot from being a moment to being a policy,” meaning your agent spends within a budget and a set of rules you define up front, instead of approving each transaction one by one.
