Google adds Tool Retrieval to let AI agents manage hundreds of tools

Technology
November 21, 2026 · 8:00 AM3 min read

Google has introduced Tool Retrieval, a new preview feature in the Gemini API that lets artificial intelligence agents tap into hundreds of software tools without loading every tool’s full definition into each request.

The feature, announced on the Google blog, is designed to cut input token usage and latency as developers build agents with increasingly large tool libraries.

What the feature does

Tool Retrieval is now in preview in the Gemini Interactions API , letting developers load tool parameter schemas on demand instead of sending their full tool library in every prompt.

It is designed to reduce input tokens and keep tool selection accurate as an agent scales. Google said when an agent connects to dozens or even hundreds of tools, sending every full parameter schema on every request adds up fast.

Tool Retrieval lets your agent keep its full tool library on standby and load parameter schemas only when the current task needs them.

Developers can turn the feature on by adding a tool_search tool and setting defer_loading to true in their requests. Once that is configured, when the model needs a tool, it asks for it by name, the API loads the full schema, and the model calls the function , all inside the same request.

Why large tool libraries get expensive

Before this change, agents wired into large sets of functions faced a growing overhead problem. Google says loading everything upfront increases input tokens, slows responses and makes tool selection less accurate.

That tradeoff matters for companies building complex agents similar to the approach covered in GCN’s report on GitHub’s Agentic Workflows tool, where automated systems must juggle many possible actions at once.

Tool Retrieval addresses the problem by splitting tool information into two tiers. It exposes the name and description of each deferred function while keeping its parameter schema out of the initial context.

After Gemini selects a function, the API loads its schema by name and emits the function call, with server-side retrieval completing that sequence within one request.

Server and client options

Google is offering two ways to run the retrieval step. There is no index to build or host. Gemini runs the search on the server, so the only code changes are the new tool and a flag on the tools you defer.

For companies that maintain their own systems, developers who maintain their own tool registry, embedding index or per-user permission filter can use client-side retrieval by declaring tool_search with “execution”: “client” and defining their search parameters.

That client-side option carries a tradeoff. Client-side retrieval adds one round trip per search, because your app runs the lookup.

But the client-side option does not require tools to be declared upfront , so new tools can be exposed mid-conversation when a user connects an integration.

Pricing and model support

Google said the feature does not add a new fee structure. Tool Retrieval has no separate charge , with input and output tokens billed at standard rates for the model called.

The capability is limited to Google’s newer model lineup, as Tool Retrieval works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API.

That includes Gemini 4 Hydrogen, the coding focused model GCN covered in its report on Gemini 4 Hydrogen’s launch.

The rollout also extends to external integrations. Google confirmed Tool Retrieval also works with remote MCP servers in the Gemini API , meaning agents connected to outside services through the Model Context Protocol can benefit from the same deferred loading approach.

Because the feature remains in preview, Google cautioned that it carries some risk. One outlet summarizing the documentation noted that any system that decides which tools to surface introduces its own failure mode: if the retrieval step misses the right tool, the model never gets a chance to pick it.

Google has not stated when the feature will move to general availability.

Hugo Rojas
Editor
Hugo Rojas is the editor of GCN. With a Master of Science in Engineering, he specializes in technology, data, and science, and brings a human-centered perspective informed by psychology.