Since co-developing the Agentic Commerce Protocol with OpenAI in September 2025, Stripe has shipped four protocol releases and onboarded leading brands. Their honest retrospective covers catalog fragmentation, real-time data demands, scoped tokens, and why sellers can't bet on a single protocol.
ACP co-developed by Stripe and OpenAI, September 2025. Four releases as of March 2026. Data from Stripe blog retrospective.
ACP launched with OpenAI. Basic product discovery, add-to-cart, checkout endpoints. RESTful interface.
Payment handlers + scoped tokens added. Agents can authorize payments with buyer-delegated credentials scoped to specific transactions.
Extensions system launched (starting with discounts). Sellers can expose promotional logic to agents. Built-in buyer auth added.
Native MCP transport added. Agents using Claude, GPT, and others can interact with ACP sellers via the Model Context Protocol natively.
The most surprising finding from Stripe's first generation of agentic commerce was the cost of catalog management. Different AI agents — ChatGPT, Claude, Google's shopping agents, Perplexity — want product data in different formats, with different field names, different image size requirements, different attribute structures. Before Stripe's unified suite, some brands were maintaining six separate product catalog exports just to be listed across the major AI agents. Every time an agent updated its required format, each brand had to rebuild and re-sync their integration.
This fragmentation is not a temporary growing pain. It reflects the reality that each AI agent is optimized differently: some prioritize structured attributes for filtering, others need rich natural language descriptions for embedding, others require real-time availability signals in the feed itself. Stripe's solution is a single canonical catalog uploaded once that the Agentic Commerce Suite translates into each agent's required format automatically — handling the syndication layer that brands shouldn't have to build themselves.
A recurring request from AI platform partners was about inventory verification: agents need to confirm products are in stock not at the time the catalog was last updated, but right now, at the moment the agent is about to initiate a purchase. Some platforms asked whether inventory verification could happen down to the millisecond. This drove Stripe to build dedicated real-time availability APIs into the Agentic Commerce Suite rather than relying on periodic catalog syncs.
The addition of scoped tokens in v1.1 addressed the authorization problem that had emerged early in live deployments. When a user authorizes an AI agent to purchase on their behalf, that authorization should be scoped to the specific transaction — not to the user's full account. Scoped tokens let a buyer say "authorize this agent to spend up to $200 on running shoes in the next 30 minutes" rather than granting the agent a permanent, broad credential. The addition of native MCP transport in v1.3 reflects how quickly the developer ecosystem has shifted: most agent frameworks now have first-class MCP support, and ACP sellers need to be accessible via that transport layer without requiring separate integrations.
The protocol-agnosticism lesson is perhaps the most important: by the time Stripe shipped ACP, Google had launched its own Universal Commerce Protocol (UCP). The Agentic Commerce Suite was built from the start as a protocol translation layer precisely because Stripe anticipated this fragmentation. Sellers using the suite don't need to bet their roadmap on any single spec — they're covered across ACP, Google UCP, and whatever comes next.
Different AI agents want different catalog formats. Without a unified layer, brands end up maintaining 6+ separate exports. One canonical feed, multiple outputs.
Agents need to verify stock availability at the moment of purchase decision, not based on a batch catalog export from hours ago. Stale data creates failed transactions.
Buyer-delegated purchases need transaction-scoped authorization. Full account credentials in agent hands create unacceptable risk. Scoping is a trust prerequisite.
ACP and Google UCP are both live. More will follow. Sellers who build natively to one protocol will rebuild constantly. Protocol-agnostic layers absorb the churn.
The Model Context Protocol is how most agent frameworks now communicate. ACP sellers need native MCP transport to be accessible from Claude, GPT, and others without friction.
Promotional logic — coupons, bundle deals, loyalty discounts — needs to be exposed via structured extensions so agents can apply them correctly during the purchase flow.
Human checkout flows (CAPTCHAs, manual address entry, UI confirmation dialogs) break agent workflows by design. Programmatic checkout APIs are the right abstraction.
Agents can't transact with sellers they can't find. robots.txt, structured product feeds, and agent-friendly sitemaps are pre-requisites, not optional enhancements.
Requiring agents to navigate OAuth login flows before purchasing creates drop-off. Native buyer auth in ACP v1.2 allows agents to authenticate on behalf of the buyer within the protocol flow itself.
Discovery + checkout + order status + returns must all be agent-accessible. A seller who implements checkout but not order status creates unresolvable agent states when fulfillment questions arise.
Brands reported maintaining up to 6 separate catalog formats before Stripe's unified suite. Source: Stripe blog retrospective (2026).
Stripe. (2026). 10 things we learned building for the first generation of agentic commerce. Stripe Blog. stripe.com/blog/10-lessons