Most agentic commerce talk is about agents doing things for you. The more important question is who approves what the agent does. Shopify Sidekick app extensions answer it with a clear pattern: agents propose, merchants approve.
This is the part of Spring '26 that gets less attention than UCP and the framework-agnostic Hydrogen preview, but it quietly defines how AI should touch a live storefront.
What Sidekick app extensions actually are
Sidekick is the AI assistant built into Shopify admin. App extensions let your app plug into it.
Shopify describes three capabilities for app developers.
First, data extensions. Using admin.app.tools.data, your app can expose read-only tools that Sidekick calls inside the sandbox. A merchant can ask a question and Sidekick answers using your app's data, returning structured JSON and resource links.
Second, deep-link actions. Using admin.app.intent.link, Sidekick can send a merchant straight to the right page in your app. Using admin.app.intent.render, it can render a focused admin UI surface. These are navigation and focus, not mutation.
Third, scoped action extensions. Sidekick can suggest a change and bring up the right UI, but the merchant stays in control of what gets updated. Shopify's own words: merchants stay in control.
That last line is the whole design philosophy.
The propose-then-approve model
Here is the pattern that matters.
An agent should not silently write to your live store. It should stage a change in a real, visible surface, then wait for a human to confirm.
In practice that looks like this. The merchant asks Sidekick to do something. Sidekick calls a read-only tool to understand the current state. It proposes a specific change. It opens the relevant editing surface inside your app with that change staged. The merchant reviews it and clicks Save. Only then does the app commit.
No headless writes. No invisible mutations. No agent reaching directly into production data.
This is not a limitation. It is the correct default for anything that affects what shoppers see.
Why this is the right default for storefront editing
Storefronts are brand surfaces. A bad automated edit is not a small bug. It is a live mistake in front of customers.
The propose-then-approve model gives you three things that pure automation does not.
A clear audit trail. Every committed change passed through a human click, so you know who approved what.
A safe blast radius. Staged changes are previewable and reversible before they go live, instead of after a customer complaint.
Trust that scales. Merchants will let agents do more once they trust that nothing ships without a confirm step. Remove the confirm step and you lose the trust that makes agent adoption possible.
This is also why read-only and deep-link capabilities should come first. An agent that can find the right thing and take you to it is immediately useful and carries almost no risk. Editing rights come later, gated behind explicit confirmation.
What this means for Hydrogen and CMS teams
If you build storefronts, themes, or a visual editing layer on top of Shopify, this pattern should shape your roadmap.
Start with read-only data tools. Let an agent answer questions like which pages changed recently, which page is missing SEO data, or which template a resource uses. That alone removes a lot of manual digging.
Add deep-link actions next. Let Sidekick open a specific project, page, or template in your editor. Navigation is safe and high value.
Treat editing as the last and most careful step. When an agent suggests a content or layout change, stage it inside your editor, show the merchant exactly what will change, and commit only on Save.
This is the difference between an agent that edits your store and an agent that helps a merchant edit their store. The second one is the product people will actually adopt.
How this connects to agent-ready storefronts
A storefront is not agent-ready just because a coding assistant can edit its files. Agent-readiness is about clear, safe contracts.
That means standard storefront events and actions so apps and agents can read and update cart and storefront state reliably. It means clean product, variant, price, and availability data. And it means a merchant editing layer where agent proposals are staged and confirmed, not applied blindly.
The Sidekick model and the framework-agnostic Hydrogen direction point the same way. Shopify wants its commerce primitives to be usable by agents across surfaces, while keeping a human in the loop for anything that changes the live experience.
Bottom line
The headline feature of agentic commerce is not that agents can act. It is that they can act safely.
Shopify Sidekick app extensions encode that as a default: agents search, agents suggest, agents open the right surface, and merchants approve before anything ships.
If you are building storefront or CMS tooling on Shopify, adopt the same order. Read-only first. Deep links second. Staged, merchant-confirmed edits last. Propose, then approve.
