According to Google's announcement, connecting an agent to dozens or hundreds of tools creates overhead when every full parameter schema is sent on every request. Google says this inflates input token counts, increases latency, and lowers tool-selection accuracy as the list of candidate tools grows. With Tool Retrieval, an agent keeps its library on standby and pulls in parameter schemas only when the current task requires them.1

Google explains that the model still sees the name and description of every deferred tool, while the parameter schemas stay out of context. In Google's refund example, a user asks for a 40-dollar refund on order A-1042 while over 50 other functions remain deferred. Inside a single API call, the model searches for issue_refund, the API loads its schema, and the model emits the function call. All three steps appear in interaction.steps so developers can inspect which tools the model considered.1

Google notes that Gemini runs server-side search automatically, so developers do not need to build or host an index. Teams can keep frequently used tools loaded upfront and defer the long tail. The post adds that Tool Retrieval works with prefix caching because deferred schemas stay out of the prompt's start, keeping the cached prefix unchanged across turns. Developers can also set defer_loading to true on a remote mcp_server tool to load every tool that server exposes on demand.1

For teams with a custom tool registry, embedding index, or per-user permission filter, Google supports client-side retrieval by declaring tool_search with execution set to client. According to the post, this option lets applications expose new tools mid-conversation without declaring them upfront, such as when a user connects an integration. The tradeoff is that client-side retrieval adds one round trip per search because the application runs the lookup. Google also notes that turning on tool_search adds a small, fixed block of input tokens.1