
Core announcement and preview API scope
Google has introduced Tool Retrieval in preview inside the Gemini Interactions API, giving agent developers a built-in mechanism to load function parameter schemas on demand rather than transmitting an entire tool catalog on every turn. Detailed on November 19, 2026 by Gemini API Product Manager Lucia Loher and Google DeepMind Research Scientist Nora Kassner in the official Google AI announcement, the capability targets agents connected to dozens or hundreds of software tools. Under conventional function-calling setups, sending every parameter schema in every prompt inflates input token counts, increases latency, and degrades tool-selection accuracy as candidate lists expand. Tool Retrieval keeps a library on standby and pulls detailed schemas into context only when a task requires them. The preview supports Gemini 3.6 and later models—including the gemini-4-hydrogen model shown in Google's code sample—via client.interactions.create.
Server-side and client-side retrieval architecture
To enable server-side retrieval, developers include {"type": "tool_search"} in the request's tools array and set defer_loading: true on long-tail functions while leaving frequently invoked tools undeferred upfront. Even when deferred, each tool's name and description remain visible to the model so it knows which actions exist without carrying full parameter definitions in context. When a prompt requires a deferred function—such as an issue_refund call in Google's order-damage example—the model searches for it by name, the Gemini API loads the schema on Google's servers with no developer-built or hosted index required, and the model emits the function_call within the same API call.
Google exposes this internal sequence in interaction.steps using three distinct record types: tool_search_call (listing requested function_names), tool_search_result (returning the loaded function schema), and function_call (carrying execution arguments), alongside id, call_id, and signature fields before the host application returns a normal function_result. For architectures that already rely on a custom tool registry, an embedding index, or per-user permission filters, developers can instead configure tool_search with "execution": "client" and custom search parameters. Client-side retrieval requires no upfront tool declarations, enabling applications to expose newly connected integrations mid-conversation by returning matching {"type": "function", ...} definitions, though each search adds one network round trip.
Prefix caching, MCP servers, and workflow impact
For engineering teams operating multi-turn agents, Google highlights two integration benefits designed to simplify production scaling. First, Tool Retrieval preserves turn-to-turn prefix caching: deferred parameter schemas stay out of the beginning of the prompt, and any schemas retrieved during execution are appended later in the conversation so the cached prefix remains unchanged across turns. Second, the same on-demand loading model extends to remote MCP servers in the Gemini API. Setting defer_loading: true on an mcp_server tool definition loads every tool exposed by that server dynamically, allowing developers to connect large remote servers without enumerating each endpoint individually. Together with the step-level audit trail inside interaction.steps, these controls let developers consolidate broad toolsets into one agent loop and debug ambiguous tool names or descriptions without maintaining separate external routing layers.
Operational caveats, attributed claims, and governance limits
Several practical boundaries shape how teams should evaluate the release. Tool Retrieval is currently in preview, restricted to the Gemini Interactions API, and requires Gemini 3.6 or later models, which means existing integrations on older Gemini models or generic multi-provider wrappers cannot use it without API migration. While Google attributes lower input token counts, reduced latency, and higher tool-selection accuracy to on-demand schema loading, the announcement provides no empirical benchmarks, latency percentiles, or pricing metrics; teams should treat those gains as vendor claims until verified on their own workloads. In server-side mode, deferred tools still keep their names and descriptions in context, so oversized catalogs with verbose descriptions can still create token overhead or naming confusion, whereas client-side execution trades an extra round trip for custom permission filtering. Finally, the announcement notes no licensing changes or updated preview SLA terms, leaving enterprise teams to weigh preview stability and platform coupling before production rollout.
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.



