AgentsAutonomy & tool use
Google previews Tool Retrieval in the Gemini API, letting agents load tool schemas on demand
Google's Gemini Interactions API now lets agents keep only tool names and descriptions in context, fetching full parameter schemas when needed. The preview, for Gemini 3.6 and later, matches deferred-loading features Anthropic and OpenAI already offer.
From the Google AI article of November 19, 2026 and cited sources: Developers add a tool_search tool and mark tools with defer_loading. [1]
SOURCES[1] Introducing Tool Retrieval in the Gemini API (Google) blog.google
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Developers add a tool_search tool and mark tools with defer_loading. [1]
- Original source
- Google AI article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 19, 2026 article published
- Covered
- Nov 21 edition
- Evidence
- 7 sources across 7 sites
- 1Introducing Tool Retrieval in the Gemini API (Google)blog.google
- 2Google shrinks idle tools in Gemini agents to about 26…ppc.land
- 3Introducing advanced tool use (Anthropic Engineering)anthropic.com
- 4Tool search (OpenAI docs)developers.openai.com
- 5Gemini's Tool Retrieval loads an agent's tools only when…i-scoop.eu
- 6Tool search tool (Claude docs)platform.claude.com
Reach: not measured yet

Developers add a tool_search tool and mark tools with defer_loading. The model then sees only each deferred tool's name and description. When it needs a tool, it asks for that tool's schema by name and then makes the call. Each step appears as a typed entry in the interaction log. Search runs on Google's servers by default, and a client-run mode is available for custom registries or permission filters. Setting defer_loading on an MCP server defers all of that server's tools. [1]
The design closely mirrors Anthropic's Tool Search Tool, launched in November 2025, and OpenAI's tool_search in the Responses API. Parameter names, hosted and client-run modes, and MCP-level deferral line up across the three. This makes on-demand tool loading a common pattern across vendors and eases migration between them, although the three APIs are not identical. [1] [3] [4] [6]
The only Gemini-specific figure is about 26 tokens per deferred tool, which comes from Google's own early testing. Savings depend on how many tools an agent has. Third-party estimates suggest deferral can slightly increase input for very small toolsets. Prompt caching also shrinks the cost gap, and each schema the model loads is billed as input for the rest of the conversation. The feature carries no surcharge beyond standard token rates. [2] [5]
the mechanism and pricing are documented, but there are no published Gemini benchmarks. Implication: agents with large or growing tool catalogs, especially big MCP servers, may cut context costs.
Read the full assessment
Small, stable toolsets may still do better with cached up-front loading.
Executive brief
Google has put Tool Retrieval into preview in the Gemini Interactions API. Agents can now leave most tool parameter schemas out of the prompt and load each one only when it's needed. The surprising part is what's missing: Google gives no savings or accuracy figures for Gemini. The only number comes from Google's documentation as reported by PPC Land: about 26 tokens per deferred tool, from early testing. The design is very close to features Anthropic and OpenAI already ship. That makes on-demand tool loading a common pattern across vendors, not something unique to Google. There's no separate charge, but each schema the model loads is billed as input for the rest of the conversation.
What changed and event timeline
Anthropic ships Tool Search Tool
Claude gets
defer_loadingand on-demand tool discovery. Anthropic says tool-definition tokens fell from about 77K to about 8.7K in its tests ().OpenAI adds
tool_searchThe Responses API supports hosted and client-run tool search for
gpt-5.4and later models (). Secondary coverage cites a 47% token reduction across 250 tasks ().Google previews Tool Retrieval
Developers add a
tool_searchtool and mark tools withdefer_loading: true. It works on Gemini 3.6 and later in the Interactions API ().First cost analyses appear
i-scoop works through break-even scenarios and argues that teams with fewer than about 20 stable tools should keep loading them up front ().
Capabilities and access
Read the full section
- Models: Gemini 3.6 and later, including all Gemini 4 models. The code samples use
gemini-4-hydrogen(Google). - Surface: Interactions API only, in preview. The older
generateContentendpoint isn't supported (i-scoop). - Modes: Search runs on Google's servers by default, inside one request. A client-run mode (
"execution": "client") lets you use your own registry or permission filters. - MCP: Setting
defer_loadingon anmcp_servertool defers every tool that server exposes. - Price: No surcharge. Tokens are billed at the model's standard rates.
Technical analysis for researchers and developers
- What the model sees: It keeps every deferred tool's name and description, but not its parameter schema.
- Caching: Loaded schemas are added later in the conversation, so the cached start of the prompt stays the same between turns.
- Client mode: Each search costs one extra round trip. In return, you can add tools mid-conversation.
Read the full section
- What the model sees: It keeps every deferred tool's name and description, but not its parameter schema. When it needs a tool it asks by name, the API returns the schema, and the model emits the call. All three steps show up as typed entries in
interaction.steps(tool_search_call,tool_search_result,function_call), which helps with debugging (Google). - Caching: Loaded schemas are added later in the conversation, so the cached start of the prompt stays the same between turns. OpenAI says the same about its version (OpenAI docs).
- Client mode: Each search costs one extra round trip. In return, you can add tools mid-conversation.
- Implementation gotchas: If you leave out
tool_search_result, loaded schemas are silently dropped. Interactions are stored for 55 days on paid projects (i-scoop). - Evaluation: Google publishes no benchmark or test method.
Claims and evidence
- Fewer tokens, lower latency, better tool choice. Google says this, but publishes no Gemini numbers for any of the three (Google).
- About 26 tokens per deferred tool. This comes from Google's own early testing (PPC Land).
- 150 tools, 2 of them called: input falls from 240K to 19.8K tokens (92%). This is i-scoop's own arithmetic, not a measurement (i-scoop).
Read the full section
- Fewer tokens, lower latency, better tool choice. Google says this, but publishes no Gemini numbers for any of the three (Google).
- About 26 tokens per deferred tool. This comes from Google's own early testing (PPC Land).
- 150 tools, 2 of them called: input falls from 240K to 19.8K tokens (92%). This is i-scoop's own arithmetic, not a measurement (i-scoop).
- Better tool selection on MCP tests. Anthropic reports Opus 4 rising from 49% to 74% and Opus 4.5 from 79.5% to 88.1% with tool search. These figures are Anthropic's own and cover Claude, not Gemini (Anthropic).
- Independent corroboration: None of the reviewed sources tests Gemini Tool Retrieval itself.
Context and prior work
Tool definitions filling up the context window has been a well-known problem for agents connected to many MCP servers. Anthropic shipped a beta tool search with defer_loading in November 2025 (Anthropic docs). Google's parameter names, step types and MCP-level deferral closely match both.
Read the full section
Tool definitions filling up the context window has been a well-known problem for agents connected to many MCP servers. Anthropic shipped a beta tool search with defer_loading in November 2025 (Anthropic docs). OpenAI followed with hosted and client-run modes and namespaces (OpenAI docs). Google's parameter names, step types and MCP-level deferral closely match both. That makes moving between vendors easier, though the three APIs are not identical.
Limitations, safety and contested findings
- Small libraries: Deferral can cost more. With five tools, i-scoop estimates input rises from 8,000 to 8,100 tokens before any searches (i-scoop).
- Caching narrows the gap: On Gemini 4 Hydrogen, i-scoop estimates about $0.10 per conversation with cached up-front schemas versus $0.03 deferred, a much smaller gap than raw token counts suggest.
- Search has its own cost: Each server-side search adds an extra model pass, and loaded schemas stay billed for the rest of the conversation (PPC Land).
Read the full section
- Small libraries: Deferral can cost more. With five tools, i-scoop estimates input rises from 8,000 to 8,100 tokens before any searches (i-scoop).
- Caching narrows the gap: On Gemini 4 Hydrogen, i-scoop estimates about $0.10 per conversation with cached up-front schemas versus $0.03 deferred, a much smaller gap than raw token counts suggest.
- Search has its own cost: Each server-side search adds an extra model pass, and loaded schemas stay billed for the rest of the conversation (PPC Land).
- Descriptions decide selection: Vague or overlapping tool descriptions lead to wrong picks. Whether the feature is available in EU regions is unclear (i-scoop).
- Preview stability: OpenAI's community forum has bug reports of HTTP 500 errors in some
tool_searchsetups, which shows the pattern is still maturing (OpenAI forum).
Business and practitioner implications
- Who benefits: Agents with large or fast-growing tool catalogs, especially big MCP servers.
- Test before switching: Run it on your own traffic and track tokens, latency and tool-selection errors. Google also tells developers to measure it themselves.
- Treat descriptions as a product surface: Tool names and descriptions now drive selection, so they need review like an internal API.
Read the full section
- Who benefits: Agents with large or fast-growing tool catalogs, especially big MCP servers. For small, stable toolsets, static loading plus caching may still be cheaper.
- Test before switching: Run it on your own traffic and track tokens, latency and tool-selection errors. Google also tells developers to measure it themselves.
- Treat descriptions as a product surface: Tool names and descriptions now drive selection, so they need review like an internal API.
- Platform lock-in: It requires the Interactions API, which means migrating off
generateContent. - Fewer reasons to stay with one vendor: All three major vendors now offer similar deferred loading, so this feature alone won't decide a vendor choice.