Friday, 20 November 2026, story 1 of 7
Google adds Tool Retrieval to Gemini API preview for large tool sets
Agents can defer parameter schemas and load them on demand. Google says this cuts input tokens and latency but gives no figures.

What happened
Google has put Tool Retrieval into preview in the Gemini Interactions API, according to a Nov 19 post. Developers add a tool_search tool and set defer_loading to true on the tools to hold back. The model sees only each deferred tool's name and description. When it needs one, the API loads the full parameter schema inside the same request.
Google says this reduces input tokens and latency and keeps tool selection accurate as libraries grow. The post gives no measurements. It works with Gemini 3.6 and later models, including all Gemini 4 models, and with remote MCP servers. Google says there is no separate charge.
What it means
The saving is not guaranteed. Google's pricing note says each loaded schema is billed as input for the rest of the conversation, and tool_search adds a small fixed block of tokens. Teams with few tools, or tools used on nearly every turn, may gain little. Google itself says to keep frequent tools undeferred and to measure against your own traffic.
The design also puts weight on tool names and descriptions, because those are all the model sees before it requests a schema. Vague descriptions may mean the right tool is never requested. The steps log shows which tools the model searched for, which helps with checking. Client-side search suits teams with their own registry or permission filter, at one round trip per search. The post gives no date for general availability.
What to do: If your agent carries dozens of tools, test deferring the rarely used ones on real traffic and compare billed input tokens and tool-selection errors before and after.
Source
Filed under Agents. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.