AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

Gemini API can now load tool parameter schemas only when the model asks for them

AI-generated contentWritten by AI

blog.google, latent.space, storage.googleapis.com, techcrunch.com · 5 sources · Sources published: 18–19 November 2026 · Some source publication dates are unknown

Tool Retrieval, which Google has released in preview, keeps the parameter schemas of an agent’s tools out of the prompt until the model needs them. Google says this cuts token use with large tool libraries, but its blog post presents no measurements.

On 19 November Google added Tool Retrieval in preview to the Gemini API’s Interactions API. The feature is designed for AI agents (programs that act through a model) with tens or hundreds of tools available.

In a company blog post, Lucia Loher, a product manager for the Gemini API, and Nora Kassner, a researcher at Google DeepMind, wrote that sending every parameter schema with each request adds input tokens and increases latency. They also wrote that the model selects tools less accurately when it has many options. A parameter schema describes what data a tool accepts and in what format.

Developers mark some tools as deferred. The model still sees their names and descriptions, but not their parameter schemas. When the model needs a tool, it asks for it by name. The API loads the schema, and the model makes the tool call within the same request. In Google’s example, the model uses this method to find a refund tool among more than 50 deferred functions. The developer’s own code carries out the refund. Google runs the search on its own servers, so developers do not need to build a search index.

Google advises keeping the most frequently used tools loaded and deferring rarely needed ones. Developers can defer all the tools of a remote MCP server with a single setting.

Google does not charge separately for the feature. The search tool adds a small, fixed number of input tokens, and each loaded schema is billed as input for the rest of the conversation. Developers save tokens if the unloaded schemas contain more tokens than the loaded ones. Google presents no measured savings or accuracy figures and instead advises developers to measure the effect on their own traffic.

The cache stores the repeated beginning of the prompt so that it does not have to be processed again on every turn. Google says Tool Retrieval does not change this beginning, because deferred schemas are not at the start of the prompt and loaded ones are added to the conversation later.

Developers with their own tool registry or a per-user permission filter can run the search in their own application. Each search then needs one extra round trip, but developers can add new tools mid-conversation.

Tool Retrieval works in the Interactions API with Gemini 3.6 and later models, including all Gemini 4 models. To enable it, developers add the tool_search tool to the request and the setting defer_loading: true to the tools they want to defer. Detailed instructions are in the Gemini API documentation.

Report an error in this story