What happened
Google announced Tool Retrieval on November 19, 2026, as a preview feature in the Gemini Interactions API. The company describes it as a way to handle agents connected to dozens or hundreds of tools without sending every complete parameter schema on every request. Developers can add a tool_search tool and mark selected functions with defer_loading: true. The model still sees each deferred tool’s name and description, while its parameter schema stays out of the initial context.
When the model decides it needs a deferred tool, it requests the tool by name. The API loads that tool’s full schema and the model makes the function call within the same API request. Google’s example uses an issue_refund function and says the search call, search result and function call appear as separate entries in interaction.steps, making it possible to inspect which tools were considered. The application still runs the function and returns a function_result; Tool Retrieval does not itself execute the refund or replace application-side handling.
Google says there is no index to build or host for server-side retrieval. The code changes are adding tool_search and setting the defer flag on the tools to hold back. Developers can keep frequently used tools loaded up front and defer the long tail. Google also says deferred schemas stay outside the prompt prefix, while schemas loaded later are added farther into the conversation, allowing the cached prefix to remain the same across turns. The feature works with remote MCP servers: marking an mcp_server tool as deferred loads the server’s exposed tools on demand, without declaring each one individually.
A client-side option is intended for teams that maintain their own tool registry, embedding index or per-user permission filter. In that mode, the developer defines a client-executed tool_search and returns matching function schemas in a follow-up interaction. Google notes that this adds one round trip per search. Client-side retrieval does not require tools to be declared up front, so an application can expose tools after a user connects an integration. The feature works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API.
Google says Tool Retrieval has no separate charge, but it is not token-free: tool_search adds a small fixed block of input tokens, and each loaded schema is billed as input for the rest of the conversation, as if declared up front. Savings occur only when the schemas that are never loaded outweigh the retrieval overhead. Google recommends measuring performance against a team’s own traffic.
What the archive adds
New Runtime’s earlier coverage on November 6, 2026, described Google’s File Search Tool as a managed retrieval system for connecting Gemini API models to a developer’s own data. (Gemini Notebook API) Tool Retrieval is a new fact on the broader story of Gemini adding managed ways to connect agents to external resources, but its focus is selecting callable tools rather than retrieving documents.
Trend signal
This supports agent protocols become the interoperability layer: the feature explicitly works with remote MCP servers and can load their tools on demand. The announcement is a concrete example of an API managing the boundary between an agent and a tool service, though it does not establish how broadly this pattern will be adopted.
Why now
Google frames the feature around a scaling problem: a growing catalog of functions increases prompt size and can make tool selection less accurate. The release is a response to agents that must work across dozens or hundreds of tools, while keeping frequently used functions immediately available.
Why it matters
Large function catalogs impose a cost before an agent has chosen what to do. Deferring schemas gives teams a way to keep a broad capability set discoverable without putting every parameter definition into every request. That may reduce input tokens and latency, and Google says it is intended to help tool selection remain accurate as candidate lists grow.
The design also makes a useful distinction between a stable default set and a long tail: keep common functions loaded, and retrieve less frequent ones as needed. Teams with custom registries or permission filters can own retrieval on the client, trading control and dynamic availability for an extra network round trip. Since the feature’s token economics depend on real usage, the relevant decision is empirical: compare the fixed retrieval overhead and schemas actually loaded with the cost of declaring the full catalog upfront.
Limits and uncertainty
Tool Retrieval is in preview, and Google does not quantify token or latency reductions, provide benchmark results, or specify a maximum tool-library size in the announcement. The claimed improvement in tool-selection accuracy is not accompanied by measured results. The company’s pricing description says the search adds a small fixed input-token block but does not state its size here; loaded schemas continue to incur normal input-token charges. Actual savings therefore depend on a workload’s tool-use distribution and schema sizes. Google’s announcement also does not establish how preview behavior or limits may change before general availability.
Skill to build now
Instrument tool usage before migrating a large catalog. For one agent, record which functions are called, how often each schema is needed, and the input tokens consumed by the current setup. Then defer the long-tail functions while leaving frequent ones upfront, and compare token use and latency against the baseline. If your application enforces per-user permissions through its own registry, test client-side retrieval separately and include its added round trip in the comparison.
Reader takeaway
Tool Retrieval moves function-schema loading from a fixed prompt cost toward an on-demand cost. It is most promising when an agent has a large catalog but uses only a subset per task; measure the retrieval overhead and actual schema load rate before assuming it saves tokens.