Gemini Tool Retrieval loads agent tools on demand
Google launched Tool Retrieval in preview in the Gemini Interactions API on 19 November. It lets an agent load detailed tool schemas when needed rather than carrying every definition in each request…
Google launched Tool Retrieval in preview in the Gemini Interactions API on 19 November. It lets an agent load detailed tool schemas when needed rather than carrying every definition in each request. For commerce app teams, the relevant test is whether an expanding collection of order, inventory and customer tools can stay manageable without sacrificing reliable selection.
In Google’s announcement, developers enable tool_search and mark deferred tools with defer_loading. Their names and descriptions remain visible; parameter schemas arrive on demand. The feature also supports remote MCP servers, and client-side retrieval can consult an app’s own registry or permission filter.
Deferring a tool is not permission to execute it
Google illustrates the mechanism with an order-refund function. Retrieving its schema helps the model form a call; your application remains responsible for carrying out the operation. A schema that says an amount is a number does not establish that the shopper owns the order or that a refund is allowed.
Our recommendation is to keep merchant identity, staff permissions and approval requirements in the execution path. Do not treat a tool becoming visible as an authorisation decision. Test attempts to request another store’s order, repeat a refund or exceed the permitted amount before connecting a prototype to live records.
Measure the library you actually use
Google says Tool Retrieval has no separate fee, but normal token charges still apply. A small search overhead and the schemas that are loaded can offset savings. The preview supports Gemini 3.6 and later models, including Gemini 4.
A useful commerce evaluation would compare the same task set with full schemas and deferred schemas: order lookup, stock questions, a return request and an unusual administrative task. Measure correct tool choice, missing arguments, extra turns, latency and total input tokens. Include unsuccessful requests; averages drawn only from completed tasks can hide a costly retry path.
Start with tools your agent rarely needs, while keeping frequent operations readily available. Give similarly named tools distinct descriptions, and review the selected call before blaming the model for ambiguity. For example, finding a return policy, creating a return and issuing a refund are separate operations even when shoppers describe all three as “return this”.
Keep the preview boundary visible
This release changes how an agent discovers tool definitions. It does not itself connect a store, validate a refund policy or demonstrate a production saving for a Shopify app. Use a test environment and your own evaluation data before adopting it in a customer-facing workflow.

More from AI
See all →
