Shopify flipped agent access on for you. The only decision left is what you log.

Group the ten tools by side effect and the audit mostly writes itself. Six only read: `search_catalog`, `browse_store`, `get_product`, `show_variant`, `get_cart` and `search_shop_policies_and_faqs` ^[8][14]. Four change something a merchant answers for: `update_cart`, `cancel_cart`, `proceed_to_checkout` and `manage_orders` ^[8][14]. All of them execute inside the shopper's live session, driving the same storefront actions the theme already calls, so an agent adding an item pops the same cart drawer a human would get ^[5].

That mechanism has a second consequence nobody put in the changelog. Because the calls drive your real storefront rather than a parallel headless path ^[5], agent-driven sessions arrive in your analytics looking like ordinary ones. There is no separate bucket unless you build it.

Where the coverage stops is uneven in a way that tracks how interesting your store is. The ten are Shopify's own commerce primitives ^[4]. A product configurator, a "book a fitting" widget, a subscription manager, a B2B quote request, a quiz-to-recommendation flow, or anything shipped as a custom app or theme section is not exposed until someone registers it against `document.modelContext` ^[7]. So the free tier is thickest on the path a standard theme already made legible, and absent exactly where a merchant spent money to be different.

Registering your own tools also moves a maintenance cost onto your side of the line. The API has already been renamed once, from `navigator.modelContext` to `document.modelContext`, and Shopify carries that churn for the defaults it shipped ^[9]. It does not carry it for yours. The standard is still in a Chrome origin trial ^[3], which is the part to price: not the afternoon spent registering a configurator, but tracking a moving spec afterwards.

Then there is the thing that makes the rest unverifiable. The changelog says nothing about observability ^[10], and per the dev.to writeup most merchants who just received these tools cannot say whether a single agent has called one ^[11]. The failure mode flagged there is not an outage. It is `update_cart` returning a schema-valid response that is subtly wrong on a variant ^[13], which reaches you as a return, not an error. One log line per invocation covers it: tool name, caller, what came back. Shopify owns the tools. The record of whether they worked is yours, or it does not exist.