Google lets Gemini agents load tool instructions on demand instead of all at once
Tool Retrieval is available in preview with no separate fee. Google’s promise of lower token costs depends on how many tool schemas an agent leaves unused.
Loading page…
Tool Retrieval is available in preview with no separate fee. Google’s promise of lower token costs depends on how many tool schemas an agent leaves unused.
Listen to this story
Google’s preview feature, Tool Retrieval, lets Gemini agents keep function input schemas out of context until a task requires them. The lookup and resulting function call happen within one API request, while the developer’s application still executes the function. For agents with large tool libraries, this can reduce tokens spent on unused schemas, but retrieved schemas are billed as input for the rest of the conversation, so savings depend on which tools agents actually use.
The preview launched November 19, 2026, through the Gemini Interactions API and supports Gemini 3.6 and later, including all Gemini 4 models.
Deferred tools’ names and descriptions remain visible to Gemini; their parameter schemas are retrieved only when requested.
Google-managed retrieval needs no developer-hosted search index; client-side retrieval supports custom registries and permission filters but adds a round trip per search.
Gemini agents no longer need every tool’s detailed instructions loaded before they start a task. Google released Tool Retrieval in preview on November 19, 2026, letting agents fetch those details when needed. The feature targets large tool libraries, where Google says loading everything upfront increases input tokens, slows responses and makes tool selection less accurate.
The preview is available through the Gemini Interactions API, the interface developers use to send requests and manage interactions. It supports Gemini 3.6 and later models, including all Gemini 4 models. Developers enable managed retrieval by adding a tool_search tool and setting defer_loading: true on the functions whose details they want to hold back.
What stays visible is important: Gemini still sees each deferred tool’s name and description. What stays out of the model’s context is its parameter schema—the structured definition of the inputs that a function accepts. The model can therefore identify an available function without first receiving all the instructions needed to call it.
Google’s example uses a customer refund. Gemini requests the issue_refund function by name, the API loads its schema, and the model produces a function call containing the order identifier and refund amount. Those steps happen within one API request. The application’s own code still executes the refund and sends back the result; retrieval does not replace that execution step.
Managed retrieval requires no search index for developers to build or host. Google runs the lookup on its servers. Developers can also mix loading strategies, leaving frequently used tools available upfront while deferring less-used functions. Search, schema retrieval and the resulting function call appear in interaction.steps, exposing the sequence for debugging tool names and descriptions.
For developers with existing infrastructure, Google offers a client-side alternative. The two approaches put the lookup in different places:
The client-side option does not require tools to be declared upfront. An application can expose new ones during a conversation, such as when a user connects an integration. Separately, managed retrieval supports remote Model Context Protocol, or MCP, servers: setting the deferral flag on a server makes its tools load on demand without listing each function individually.
Tool Retrieval has no separate charge, but it is not token-free. Google’s pricing explanation says standard model input and output rates apply. Enabling tool_search adds a small, fixed block of input tokens. Each schema retrieved is then billed as input for the rest of the conversation, just like a function definition supplied at the start.
The feature also works with prefix caching, which reuses the unchanged beginning of a prompt. Deferred schemas stay out of that opening portion; schemas fetched later are added farther along in the conversation. Google says this keeps the cached prefix stable between turns, rather than changing it each time the model retrieves another tool.
Google advises developers to measure savings against their own traffic. Its stated test is whether the schemas never loaded outweigh those that are. A large library alone is not the whole calculation: which tools an agent actually retrieves determines how much schema content remains out of the conversation.
Loading discussion...
Join the conversation
Explain when custom control would be worth the extra work.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.