THE APEX TIMES
In a new preview feature, developers can keep large tool libraries on standby and load tool parameter schemas only when an agent needs them, reducing input size and improving accuracy.
Alphabet is rolling out a new capability for building AI agents in its Gemini API, aiming to make it easier for developers to connect models to very large sets of software tools without inflating prompt sizes. The company calls the feature “Tool Retrieval,” and it is available in preview in the Gemini Interactions API.
The core problem Tool Retrieval addresses is scale. When an agent can use dozens or even hundreds of tools, sending the full parameter schema for every tool on every request can quickly drive up input tokens, add latency, and make tool selection less precise as the candidate list grows. Google says the added schema volume can also complicate debugging when developers need to understand which tools the model considered.
With Tool Retrieval, the agent still knows which tools exist, but the detailed parameter schemas are held back until they are required for the current task. Google describes a model workflow in which the agent issues a lookup request by tool name, the API loads the relevant schema, and the model then proceeds to call the function. Google says all of these steps are visible in the interaction flow, specifically under, which can help developers inspect which tool names or descriptions were considered.
To turn the feature on, developers are instructed to add a tool_search tool to the request and set defer_loading: true on the specific tools they want the system to hold back. Under the described approach, the model sees the name and description of each deferred tool, so it can select from the available options, while the heavy schema content stays out of the prompt context until it is needed.
Google positions the change as a way to reduce both costs and latency for complex agent tasks. It says Tool Retrieval can lower input token counts and improve performance and accuracy as the tool library scales, because the system loads only the schemas that match the tool the model chooses for the current step. The company also says the feature is designed to keep the tool selection process accurate even when the overall number of tools is large.
The company also highlights how Tool Retrieval fits with other Gemini API capabilities. It says Tool Retrieval works with prefix caching, a technique that reuses parts of a prompt across turns to cut repeated input processing. Google’s explanation is that deferred schemas remain out of the start of the prompt and are added later when loaded, so the cached prefix can remain consistent from turn to turn.
For developers dealing with external integrations, Google says Tool Retrieval works with remote MCP servers (Model Context Protocol servers) in the Gemini API. In that setup, developers can mark an mcp_server tool with defer_loading: true so that the tools exposed by the remote server are loaded on demand, rather than requiring a full list of exposed tools to be enumerated ahead of time.
Google says Tool Retrieval does not require building or hosting a separate index. It describes Gemini doing the search on the server side, so “the only changes” are the addition of the tool_search tool and the defer_loading flag on deferred tools. Google also notes that Tool Retrieval has no separate charge beyond standard token billing for the selected model. It says turning on tool_search adds a small, fixed block of input tokens, and each schema loaded is billed as input for the rest of the conversation, the same way a tool declared up front would be. The implication is that the feature’s savings depend on whether the tool schemas you do not load outweigh the schemas you eventually retrieve.
As to availability, Google states that Tool Retrieval works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API. It also says developers can try the feature immediately in preview by following the described Interactions API request pattern, without additional infrastructure for retrieval.
Even with Google’s detailed product description, some operational questions remain for developers. The company does not provide benchmarks for latency or token savings across different tool-library sizes, nor does it specify how often retrieval is triggered relative to tool availability in typical agent conversations. Developers will likely need to test how many schema loads occur in real workloads, and how the added lookup step affects end-to-end response time.
The outlet’s screening for lower-priced “blue chip” stocks highlighted Salesforce (NYSE: CRM) as the 6th name in its ranking, framing the CRM software leader as part of a steadier large-cap technology basket.
A reported FedStart-linked achievement for OLIGO Security illustrates the practical steps needed to move AI systems into the U.S. federal environment, where authorization requirements can make or break timelines.
Intel stock has climbed about 56% over the past year, but a recent pullback has renewed questions about whether the market is pricing the company’s long-term ability to generate cash correctly, according to a Yahoo Finance analysis tied to fresh chip-focused news.
A fresh market note revisits a central question for Intel shareholders: how “factory cash” measures cash generation against heavy capital spending, and what that balance could mean for future funding needs and investor expectations.
Worldpanel by Numerator data cited by Yahoo Finance shows a +51 net score for recommending Netflix, compared with +53 a year earlier.
A recent U.S. ethics disclosure described trades that included Nvidia (NVDA) purchases valued between $152,004 and $380,000 and a later sale of up to $15,000 on Sept. 24, alongside remarks about Nvidia CEO Jensen Huang.
The U.S. Department of Commerce has designated Salesforce to help coordinate an American Super Intelligence export effort aimed at promoting priority global shipments tied to “trusted AI,” according to a report published by Yahoo Finance.
A Yahoo Finance article flags potential issues in how AMD manages and maintains key relationships outside the United States, raising questions about continuity and leverage in fast-moving Asian markets.
A Yahoo Finance report says investors are questioning the privacy implications of Meta’s AI plans after claims that an AI agent sent users messages and took actions without clear authorization. Meta has not publicly detailed the specific incident or its safeguards in the reporting.
In a study published with Science, Google Quantum AI says its qubits can sustain a quantum effect thought to power certain algorithms, and that the same tests may help expose subtle hardware errors as systems grow larger.
Comments
Community guidelines. Comments may be removed for violations.
Loading comments...