Google has put Tool Retrieval into preview in the Gemini Interactions API, a feature that stops agents from carrying every tool definition in every request. Instead of sending full parameter schemas for dozens or hundreds of functions each turn, a developer can hold most of them back and let the model pull in only the ones a task needs.

The problem is familiar to anyone who has wired an agent to a large internal toolbox. Every function definition costs input tokens on every call. Google’s announcement, written by Gemini API product manager Lucia Loher and DeepMind research scientist Nora Kassner, says a long candidate list also adds latency and makes the model worse at picking the right tool.

Setup takes two changes. A developer adds a tool_search tool to the request and marks the functions to hold back with defer_loading: true. The model still sees each deferred tool’s name and description, so it knows the option exists. When it wants one, it asks by name, the API loads that schema, and the function call happens inside the same request. Google’s demo uses a refund function. Asked to refund a damaged order, the model looks up issue_refund, gets its schema back, and emits the call in one pass.

The search runs on Google’s servers, which spares developers from building or hosting an index. Each step of the lookup shows up in interaction.steps, which gives teams a way to see which tools the model considered when a call goes wrong. That is useful for debugging vague tool descriptions, since the description is now the only thing the model has before it asks for the rest.

Teams that already keep their own registry or per-user permission filters get a second path. Declaring the search tool with client-side execution hands the lookup to the app, at the cost of one extra round trip per search. That route also allows tools to appear mid-conversation, for example when a user connects a new integration.

Remote MCP servers are covered too. Setting the deferral flag on an mcp_server tool defers everything that server exposes, which matters because large MCP servers are where tool lists tend to balloon without anyone counting.

Google also says the design cooperates with prefix caching. Held-back schemas never enter the start of the prompt, and loaded ones are appended later, so the cached prefix is stable across turns. Because loaded schemas then count as input for the rest of the conversation, an agent that touches many tools in one session will still pay for them.

Pricing follows the same logic. Google lists no separate charge for Tool Retrieval; usage is billed at the model’s standard input and output rates, plus a small fixed block of tokens for turning the search tool on. The savings arrive only when the schemas an agent never loads outweigh the ones it does, and Google itself tells developers to measure that against their own traffic.

That advice is the notable gap. The announcement includes no benchmark, no token reduction figure, and no accuracy comparison, even though lower token counts and better tool selection are the stated goals. Support is limited to the Interactions API, on Gemini 3.6 and newer, which takes in every Gemini 4 model.

For any team running an agent with more than a few dozen tools on Gemini, the next step is a controlled test: replay a day of real traffic with and without deferral, and compare token bills and wrong-tool errors before moving the tool library over.

Reported by Google (blog.google) on November 19, 2026.