Google targets a scaling problem for AI agents
Google has introduced Tool Retrieval in preview for the Gemini Interactions API, a feature intended to help AI agents work with large collections of software tools without sending every tool definition in every request.
The premise is straightforward: an agent may have access to dozens or hundreds of tools, but most individual tasks need only a small subset. Supplying every full parameter schema on every request can increase input-token use, add latency and make tool selection harder as the available list expands. Tool Retrieval is designed to keep a full tool library available while loading detailed schemas only when a task calls for them.
The feature is available in preview in the Interactions API and is designed for Gemini 3.6 and later models, according to Google’s announcement. Developers can enable it by adding a tool-search capability to a request and setting deferred loading to true for the tool definitions.
What changes for developers
Tool calling lets a model request that an application perform an action, such as querying a database, checking an internal service, retrieving a document or initiating a workflow. To use a tool correctly, the model ordinarily needs a description of the tool and a schema specifying its inputs. As an agent’s tool catalog grows, including all of that information in the prompt becomes increasingly expensive and unwieldy.
Google’s new approach separates discovery from detailed invocation. Rather than placing every parameter schema in the context up front, a developer can leave tools in a library and allow the agent to retrieve the relevant schema when it identifies a possible match for the task.
That can reduce the amount of data sent to the model at the beginning of a task. It also changes the design of an agent application: tool metadata can be treated less like a fixed, always-present prompt attachment and more like information the model can consult during its work.
Why schema loading matters
A tool schema is not merely a label. It can include a tool’s name, description, required and optional arguments, expected types, and other constraints. Large organizations may maintain substantial catalogs of internal services, while more capable agents may combine search, data systems, communications tools, scheduling systems and specialized business functions.
In that setting, the full set of schemas can consume a meaningful share of the context supplied to a model. More prompt content can increase input-token costs and, depending on the system, contribute to slower responses. Google says Tool Retrieval is intended to reduce both input token counts and latency for complex agent tasks.
Google also says the feature can improve tool-selection accuracy as tool libraries scale. That is a practical concern: a model presented with many similarly described tools may choose a less suitable one or spend more effort sorting through options. Retrieving details on demand creates a staged process in which an agent can first search its available tools and then inspect the exact parameters of the most relevant candidate.
A retrieval layer for actions
The announcement resembles the broader pattern of retrieval used in AI systems. In a retrieval-augmented workflow, an application does not put every document in a prompt; it searches a knowledge collection and supplies the most relevant material. Tool Retrieval applies a related idea to actions and interfaces rather than documents.
The analogy has limits. A tool definition can affect what an agent is permitted to do, and an incorrectly selected tool may produce an invalid operation or trigger an unwanted workflow. The feature does not remove the need for developers to define permissions, validate inputs, control which tools an agent can access and confirm consequential actions.
Google’s announcement describes Tool Retrieval as a way to load parameter schemas only when needed. It does not say that access controls, authentication or execution safeguards are handled automatically by the feature. Those remain responsibilities for the system that exposes the tools.
Preview status is significant
Because Tool Retrieval is in preview, developers should treat the release as an opportunity for evaluation rather than as a guarantee of stable, production-ready behavior. Preview features can change in their interface, supported models, performance characteristics or usage terms before broad availability.
The announcement identifies the Gemini Interactions API as the initial home for the capability. Developers already building agents with large tool inventories can test whether deferred schema loading reduces their actual prompt sizes and response times, while checking whether the retrieval step changes how reliably their agents choose the right tool.
Those measurements will matter more than the feature’s conceptual appeal. A tool library may be large in theory but compact in practice; another may have schemas so detailed that selective loading delivers a clear benefit. The best evaluation will compare representative tasks, token use, latency, selection behavior and error handling.
What the launch signals
Google’s release highlights a broader shift in agent development. The challenge is no longer just giving a model one or two functions. Developers increasingly want systems that can work across broad sets of tools while remaining responsive and manageable.
That creates pressure to make context more selective. Every available tool need not be fully described in every interaction, just as every document in an organization need not be attached to every question. Tool Retrieval offers one approach: keep the catalog accessible, retrieve the details when needed and reserve the main prompt for the current task.
For now, the announcement is a preview of that approach in Gemini’s API. Its practical value will depend on whether developers can reduce prompt overhead without losing the reliability and transparency they need when an AI system chooses and invokes software tools.
This article is based on reporting by Google AI Blog. Read the original article.
Originally published on blog.google








