DE EN
Agentic Punks
Zurück zur Übersicht

OpenAI's Superapp is not a retreat - it's an operating system for agents

23. März 2026 | Roman Zenner
Teilen:
OpenAI's Superapp is not a retreat - it's an operating system for agents

OpenAI has abandoned its checkout feature. Barely more than a few dozen of the millions of Shopify merchants have ever sold anything via ChatGPT. Has Agentic Commerce failed? Is it curtains? Not so fast.

Because while the industry is still rehearsing the swan song for OpenAI’s commerce ambitions, the company is quietly building something else: a desktop superapp that merges ChatGPT, the coding agent Codex, and the AI browser Atlas into a single application. Anyone who thinks this is just a simplification is underestimating what’s happening here.

What’s behind the super app?

As The Verge reports, Fidji Simo, OpenAI’s CEO of Applications, set the direction in an internal memo: Fragmentation into separate apps has slowed the company down and diluted quality. “Side projects” that don’t contribute to coding and business customers will be deprioritized.

The message is clear: OpenAI isn’t pulling out of commerce—it’s pulling out of the wrong layer.

The Three-Layer Problem

To understand Agentic Commerce, you need to distinguish between three layers:

Layer 1 — AI-powered research. Works today. People ask ChatGPT for purchase recommendations, compare products, and seek advice. This will be the norm in 2026.

Layer 2 — AI-assisted checkout. This is where OpenAI gave up — and that was the sensible thing to do. People research using AI interfaces, but they buy where they trust: from retailers.

Layer 3 — Autonomous agents that act independently. The future, but not yet the present.

OpenAI’s superapp targets something that lies between Layer 1 and Layer 3—and doesn’t yet have a proper name: the Agent Operating Layer. The layer where an AI agent not only makes recommendations but also navigates the web, fills out forms, compares prices, and executes workflows. Not as a store, but as an operating system.

Atlas is the key element

The most interesting component in the Superapp isn’t ChatGPT—everyone knows that. It’s Atlas, the AI-native browser. A browser built not just for people who read web pages, but also for agents that act on the web.

Combined with Codex (an agent that can write code and automate tasks) and ChatGPT as a conversational interface, this creates an architecture in which agents:

  • Read product pages and extract structured data
  • Go through checkout flows at retailers
  • Automate price comparisons across multiple stores
  • Orchestrate end-to-end business workflows

This is a fundamentally different approach than the failed attempt to integrate checkout into ChatGPT. Instead of taking the transaction away from retailers, the agent becomes a customer proxy that makes purchases on their sites.

The irony of the timing

The timing is remarkable: As OpenAI executes its superapp pivot, an Omnisend study shows that 80% of U.S. shoppers are now open to AI-assisted purchases—a jump from 34% the previous year. At the same time, 86% of respondents still express concerns, and 70% reject personalized pricing driven by AI. And Google has just expanded its Universal Commerce Protocol with new cart and catalog features designed to make shopping easier for AI agents.

So consumer acceptance is rising at the very moment OpenAI is building the infrastructure that actually enables agents to take action. No checkout button in ChatGPT, but an agent that handles the checkout process with the retailer for you.

What does this mean for retailers?

The uncomfortable truth: When agents no longer come via APIs and protocols, but through the browser—then your website suddenly becomes the most important interface again. Not for humans, but for machines. And the question is no longer “Should I sign up for Google’s UCP?”, but: Can an agent even operate my site?

OpenAI’s superapp strategy is a bet that the future of agentic commerce lies not in protocols, but in the web itself—navigated by agents that have a browser. This is less elegant than a standardized protocol. But it works from day one with any shop that has a website.

The real question is a different one: If the AI agent becomes your customers’ default browser—who then controls the shortlist?