Google announced Tool Retrieval in preview for the Gemini Interactions API on November 19. Deferred tools keep their names and descriptions visible while parameter schemas load when needed. The aim is to reduce unnecessary input context for large libraries. The announcement does not establish a universal token-saving percentage or guaranteed selection accuracy for every tool set.
The documentation requires v1beta and states that generateContent does not support the feature. It uses tool_search and defer_loading for the relevant functions. This is a particular API route with preview conditions, rather than an automatic change to every existing Gemini integration.
Discovery is not permission
Engineering teams face separate questions: which tool fits, which arguments it accepts and whether the action is authorized. Deferred loading concerns supplying a parameter description. It does not establish that a particular order, file or recipient is allowed for the operation. Those boundaries must remain in the application executing the request.
Consider our own example of two similarly named tools: reviewing a request and finally transmitting it. Their descriptions should distinguish reading from an external effect. Even selecting the correct schema is insufficient if the agent supplies the wrong identifier. Acceptance checks should detect this failure too, rather than measuring only context size.
Measuring the benefit
We suggest comparing the same realistic demonstration tasks with a consistent starting state. Record input tokens, elapsed time, correct selection and argument validity. Include an infrequently used tool and an ambiguous request. Do not select only the easiest example after seeing the outcome.
Next, test an unavailable or renamed tool. The system should produce an inspectable failure instead of executing a similar but different operation. If a person corrects the choice, include that time in the evaluation. The objective is reliable task completion, rather than simply smaller numbers in a context report.
What remains unverified
We have not run such a comparison. Google provides a mechanism and motivation; actual behavior depends on the library, descriptions and application. For a team considering deployment, a useful first outcome is a small, verifiable experiment with unchanged access rules. A preview can be useful without being presented as a final solution to every tool-selection problem.



