What happened
Google has opened a preview of Tool Retrieval in the Interactions API for Gemini, Google reported. The feature lets agents hold large tool libraries in reserve and fetch a tool's parameter schema only when a task calls for it. Developers switch it on by adding a tool_search tool to a request and marking the tools to hold back with the defer_loading flag. It is supported on Gemini 3.6 and newer models, including every Gemini 4 model.
With the flag set, the model still receives the name and description of each deferred tool but not its full schema. When it needs one, it asks for that tool by name, the API supplies the schema, and the model issues the function call, all within a single request. Each of those steps is logged in interaction.steps for debugging. Google runs the search on its servers, so developers do not have to build or host an index.
The feature also covers remote MCP servers. Setting defer_loading on an mcp_server tool makes everything that server exposes load on demand. Teams with their own tool registry or permission filters can run the search client-side instead, at the cost of one extra round trip per lookup, and can add new tools partway through a conversation. There is no extra fee. Tokens are billed at standard model rates, and enabling tool_search contributes a modest, fixed number of extra input tokens.
Context
Agents typically send every tool's full parameter schema with each request, a cost that grows as libraries expand. Google positions Tool Retrieval as compatible with prefix caching. Schemas that are deferred stay clear of the prompt's opening, while loaded ones are appended later, so the cached prefix stays stable across turns. Any schema the model loads is then charged as input tokens through the remainder of the conversation, like a tool declared up front.
Das Briefing
Die Tech-Woche in einer E-Mail.
Die lesenswerten Beiträge der Woche. Jederzeit abbestellbar.
