What changed: Google introduced Tool Retrieval in the Gemini API: by adding a tool_search tool and setting defer_loading to true on functions, the model no longer receives every tool's full parameter schema on each request, only loading the schema it needs for the current task. This cuts input token counts and latency while keeping tool-selection accuracy as a tool library grows. The feature works with Gemini 3.6 and later models, including Gemini 4 models, in the Interactions API.
Who is affected
The source refers to developer teams building agents with the Gemini API that rely on large tool libraries; no specific industry limitation is stated.
What it means for your business
For businesses building multi-tool AI agents on the Gemini API, this means they can cut input token costs and latency without sending every tool schema on each request, while keeping tool-selection accuracy as their tool count grows.
What to do
Teams running agent systems with many tools on the Gemini API should add the tool_search tool, set defer_loading to true on infrequently used functions, and measure the token and latency impact on their own traffic.
When it takes effect
The source states the feature was announced in preview in the Interactions API on Nov 19, 2026; no separate effective date is given.
Source: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-api-tool-retrieval/