What happened
Google introduced Tool , a preview capability in the Gemini Interactions API that allows developers to defer loading of tool parameter schemas until the model explicitly requests them. By adding a special `tool_search` tool and setting `defer_loading: true` on selected tools, the model sees only names and descriptions up front, fetching full schemas on demand within the same request. The feature works with Gemini 3.6 and later, including all Gemini 4 models, and integrates with remote MCP servers and client‑side registries. No separate charge is applied; schema loads are billed as regular input tokens.
Google announced the preview of Tool in the Gemini Interactions API. The new workflow requires developers to add a `tool_search` tool to their request payload and mark specific tools with `defer_loading: true`. The model retains only the name and description of deferred tools, allowing it to consider them during tool selection without the full schema present in the prompt.
When the model decides a deferred tool is needed, it issues a request for that tool’s name. The API then loads the complete parameter schema and returns it within the same interaction step, after which the model emits a `function_call`. All three steps—search, load, call—appear in the `interaction.steps` log, facilitating debugging.
The feature works with Gemini 3.6 and later models, including the Gemini 4 family, and is compatible with remote MCP (Model‑Centric Platform) servers. Developers can also implement client‑side by declaring `tool_search` with `execution: "client"`, which adds a single round‑trip for each lookup.
Billing remains unchanged: input and output tokens are charged at the standard Gemini rates. Enabling `tool_search` adds a small fixed token overhead, and each loaded schema is billed as regular input tokens. The expectation is that overall token consumption will drop when many deferred schemas are never loaded.
Why it matters
Tool addresses a practical bottleneck for developers building AI agents that need to interact with dozens or hundreds of external tools. Previously, each request had to include the full JSON schema for every tool, inflating token counts, increasing latency, and reducing the model’s ability to accurately select the right tool as the candidate list grew. By loading schemas only when required, developers can keep large tool libraries on standby without paying the token cost for unused definitions, improving efficiency and potentially lowering operating expenses. The feature also simplifies integration with large remote tool registries and supports prefix‑caching, which can further reduce latency in multi‑turn conversations. While the preview is free of extra fees, actual cost savings depend on the proportion of deferred tools that are never invoked, a factor developers will need to measure against their own traffic patterns.
Large‑scale tool libraries have become common in AI‑driven agents for e‑commerce, finance, and enterprise workflows. The need to transmit full schema definitions for every tool on each request can quickly become a cost and performance issue, especially as token pricing remains a primary expense driver for API usage.
By deferring schema loading, developers can keep a broad catalog of tools available without incurring the token cost for unused definitions. This reduces the prompt size, which directly lowers latency and improves the model’s ability to rank relevant tools accurately, as the candidate list is not cluttered with unnecessary schema details.
The preview also simplifies integration with dynamic tool registries. Developers can expose new tools mid‑conversation without pre‑declaring them, enabling more flexible, user‑driven experiences such as on‑the‑fly plugin activation or per‑user permission filters.
From a business perspective, the feature could make Gemini‑based agents more economically viable for startups and enterprises that need to support extensive tool ecosystems, potentially increasing adoption of Google’s generative AI platform.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Developers will monitor token‑usage metrics to quantify savings, and early adopters may share performance benchmarks that reveal latency improvements. Google may expand the preview to a full release, add finer‑grained control over schema loading, or integrate the capability into other Gemini API offerings. Compatibility with future Gemini model versions and any changes to pricing structures for deferred schema loads will also be important to watch.
Token‑usage analytics from early adopters to confirm projected savings versus baseline costs.
Performance benchmarks measuring latency improvements when using deferred loading versus full schema transmission.
Potential updates from Google expanding the preview to a GA (general availability) release, adding features like selective schema caching or finer‑grained billing.
Changes to pricing policies for schema loading, especially if deferred schemas become a significant portion of usage for large‑scale deployments.