Leo Hösl
/
Move Fast and Free: The Commerce Stack for the Agentic Era

ArticleMove Fast and Free: The Commerce Stack for the Agentic Era

For a decade, the default advice for a serious shop was some flavor of Shopify. Start hosted, and when you outgrow the theme editor, go headless: keep Shopify as the commerce engine, put your own frontend in front of it. I have built and maintained shops on both models, and until recently that advice was reasonable.

AI-assisted development changed the calculation, and I do not mean that as a slogan. It changed one specific variable: what a unit of custom development costs. When agents write most of the code, custom is no longer the expensive option. Rented is.

What agents actually need

An AI coding agent is only as good as what it can see and touch. Give it a codebase where the products, the checkout, the pricing logic, the CMS schema and the frontend are all code in one repository, and it can trace a feature end to end, change it, migrate the database and test the result. The whole loop is inspectable.

Now put Shopify in the middle. The checkout is a black box. The data model is fixed, extended only through metafields and apps. Rate limits sit between your code and your own data. Every capability boundary is a support article, not a source file. The agent cannot read Shopify's side of the system, so every task that crosses the boundary degrades from "change the code" to "work around the platform". You pay for that workaround in developer hours today, and you keep paying app subscriptions for capabilities that would be a module in your own backend.

Headless Shopify is the strangest deal of all in this light. You take on the full complexity of running your own frontend, and keep the platform ceiling anyway. Double the surface, same walls.

The stack that held

The stack I now reach for, and have in build for a six-market European shop with a catalog in the hundreds of thousands of products:

  • Medusa as the commerce engine. Open source, TypeScript, Postgres underneath. Products, carts, pricing, promotions and fulfillment are modules you can read, extend and replace. Custom logic is a workflow or subscriber in your repo, not an app store dependency.
  • Payload as the CMS, embedded in the same Next.js app. Content schema is code, migrations are code, editors get live preview, and per-market content (down to legally required text blocks) is editable without a deploy.
  • Next.js as the frontend, one codebase serving every market by domain, with theming and locale resolved per host.
  • Postgres and a search engine (Typesense in our case) as the data spine. Ingest pipelines, price history, review aggregates and compliance values are indexed as data, so pages read one document instead of orchestrating five APIs.

Every layer is open, local, and versioned in git. That is not ideological purity. It is the property that makes agentic development compound: each feature the agents build becomes context that makes the next feature cheaper. On a rented platform, that compounding stops at the API boundary.

The honest counterweight

Shopify still wins in specific situations. If you have no development capacity at all, sell a small catalog, and standard checkout is fine, the hosted model does what it always did, and the agentic argument barely applies because there is no codebase to compound. Fraud tooling and one-click ecosystem breadth are real assets.

But the threshold has moved, and it keeps moving. The moment your roadmap contains things the platform does not do, multi-market pricing rules, a custom quoting flow, regulated display values, deep ERP integration, the owned stack is no longer the brave choice. With agents doing the heavy lifting, it is simply the faster one, and it is yours. No rent, no ceiling, no waiting for a platform roadmap to catch up with your business.

Move fast, and own the thing you are moving.