Money and business in the Middle East.

AI

Google lets Gemini agents defer tool schemas, cutting token load

Gemini agents can now reach hundreds of tools without carrying every parameter schema in every request. Google said on November 19, 2026 that Tool Retrieval is in preview in the Gemini Interactions API, with no general availability date announced.

· Originally published by ontime+ · Last verified: 21 Nov 2026

Key Points

  1. Google opened Tool Retrieval in preview, letting Gemini agents load tool schemas only when needed.
  2. Deferred tools keep names and descriptions in context while parameter schemas stay out until requested.
  3. Developers connecting hundreds of tools or MCP servers can trim input tokens per request.

The latest:

Gemini agents can now reach hundreds of tools without carrying every parameter schema in every request. Google said on November 19, 2026 that Tool Retrieval is in preview in the Gemini Interactions API, with no general availability date announced. The company framed the problem plainly: shipping full schemas on each call piles up fast, and the feature is built to cut that load.

Details:

  • The mechanism: According to Google, developers add a tool_search tool to the request and set defer_loading: true on the tools they want held back. Each deferred tool keeps its name and description inside the model’s context, while its parameter schema stays outside until the model asks for it by name.
  • The response flow: Google’s published example runs a tool_search_call, then a tool_search_result, then a function_call, with the developer’s code returning the function_result. The interface loads the full schema and the model calls the function within the same request, according to the announcement.
  • No index required: Developers do not build or host a retrieval index of their own. Google said Gemini performs the search server-side, removing the infrastructure layer that agent builders would otherwise maintain to route a model toward the right tool.
  • Caching behavior: The feature works alongside prefix caching, according to Google. Deferred schemas stay outside the request prefix, and schemas the model loads are appended afterward, so the cached portion of the prompt is not disturbed when a tool is pulled in mid-run.
  • MCP servers: Remote MCP servers are supported in the Gemini API. Setting defer_loading: true on an mcp_server tool means every tool that server exposes is loaded on demand rather than declared upfront, according to Google — the scenario where schema volume grows fastest.
  • Client-side variant: Developers can instead declare tool_search with “execution”: “client” plus their own name, description and parameters, running the search inside their application. Google said this version adds one network round trip per search.
  • Mid-conversation discovery: Tools do not have to be declared in advance, according to Google, so new tools can surface partway through a conversation — relevant for agents whose available actions change with the task or the user’s permissions.
  • Availability: Tool Retrieval runs on Gemini 3.6 and later, including all Gemini 4 models, and only through the Interactions API. Google’s code sample uses gemini-4-hydrogen, and documentation sits at aistudio.google.com/docs/tool-retrieval. No general availability date was given.
  • The pricing: Google said there is no separate charge for Tool Retrieval; tokens bill at the model’s usual rates. Enabling tool_search adds a small fixed block of input tokens, whose size Google did not state, and each loaded schema bills as input for the rest of the conversation.
  • No benchmarks: The announcement carries no figures for token savings, latency or tool-selection accuracy, and no quotes from authors Lucia Loher, product manager for the Gemini API, and Nora Kassner, research scientist at Google DeepMind. Google’s guidance is to measure savings against actual traffic.

Between the lines:

Google’s own framing puts the burden of proof on developers: with no published savings figures and billing that charges each loaded schema as input for the rest of the conversation, the economics depend on how often an agent actually needs a given tool. The fixed token block added by tool_search, whose size Google did not disclose, sets the floor any saving must clear.

What’s next

Watch for a general availability date, disclosure of the fixed tool_search token cost, and whether Google publishes accuracy or latency numbers for server-side tool selection as preview users report results.

Read on ontime+ ↗