An agent wired to dozens or even hundreds of tools sends the schema of every one of them with every question. Even when the task needs one. Google shipped Tool Retrieval in the Gemini API: the library stays on standby, and a schema loads when the task asks for it. The feature is in preview in the Interactions API.
Fact
The facts: on 19 November 2026 Google announced Tool Retrieval in preview in the Gemini Interactions API. The idea: tool parameter schemas load on demand instead of the full library going into every prompt, designed to reduce input tokens and keep tool selection accurate as the agent scales. You turn it on by adding a tool_search tool to the request and setting defer_loading: true on the tools you want to hold back. The model still sees the name and description of every deferred tool, but the parameter schemas stay out of context. There is no index to build or host: Gemini runs the search on the server, so the only changes to your code are the new tool and a flag. It also works with remote MCP servers in the Gemini API. There is no separate charge: input and output tokens are billed at the standard rates for the model you call. It works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API. The announcement's authors: Lucia Loher, Product Manager for the Gemini API, and Nora Kassner, Research Scientist at Google DeepMind.
The model does not lose track of what it has. It sees the name and description of every deferred tool. The schema waits. When it needs it, the model asks for the tool by name, the API loads the schema and the model calls the function, all inside the same request.
The tool does not disappear. It just stops travelling with every question.
The saving is not automatic. Turning on tool_search adds a small, fixed block of input tokens, and each schema the model loads is billed as input for the rest of the conversation, the same as a tool declared up front. You save when the schemas you never load outweigh the ones you do. For a large library that is the usual case. Measure it against your own traffic.
If you keep your own tool registry, declare tool_search with execution: client and define your own search parameters. Then your app runs the lookup, and it adds one round trip per search. Tools do not need to be declared up front, so you can also expose a new tool mid-conversation, when a user connects an integration.
The test is your library, not the sample one. If the schemas you hold back outnumber the ones you load, the gain is real. If they do not, you are carrying one more flag for nothing.
Official primary sources
Google - Introducing Tool Retrieval in the Gemini API, 19.11.2026The visual is generated code-art, no third-party images.