RealityHackerOpen in RealityHacker ⇢
Tools · AI Coding Tools · published 2026-11-19T00:00:00+00:00 · via Google Blog

Gemini API Now Enables On-Demand Tool Access to Reduce Costs and Improve Efficiency

Image via Google Blog
Image via Google Blog

Google has introduced Tool Retrieval in the Gemini API, allowing agents to access hundreds of tools without loading all parameter schemas at once. This on-demand approach reduces input token usage, lowers latency, and improves tool-selection accuracy as systems scale. The feature is available in preview via the Interactions API and works with Gemini 3.6 and later models.

Expanded Detail

Google has introduced Tool Retrieval in the Gemini API, launching in preview on November 19, 2026, via the Interactions API. This new feature enables AI agents to access hundreds of tools on demand, loading only the parameter schemas of specific tools when needed, rather than including all schemas in every prompt. By setting the `defer_loading: true` flag on functions like `issue_refund`, developers can reduce input token usage, lower latency, and improve tool-selection accuracy—especially critical as agent systems scale. The feature is compatible with Gemini 3.6 and later models, and it operates seamlessly within the API’s existing function calling framework. Notably, the model retains visibility of tool names and descriptions even when schemas are deferred, enabling accurate decision-making. The system handles schema loading internally during a single API call, with steps logged in `interaction.steps` for debugging and transparency.

The development builds on prior challenges in AI agent design: as tool libraries grow, sending full schemas with every request becomes inefficient, increasing costs and slowing performance. Earlier approaches required developers to pre-load all tools or maintain external indexes, which limited scalability. Google DeepMind researchers and Gemini API product managers, including Lucia Loher and Nora Kassner, have addressed this by introducing server-side tool search, eliminating the need for client-side infrastructure. The feature also integrates with remote MCP servers, allowing on-demand loading of tools exposed by such servers without manual listing. For developers maintaining custom tool registries, client-side retrieval is supported via `execution: "client"`, enabling dynamic discovery mid-conversation. However, the long-term impact on developer workflows, especially in regulated industries, remains untested. What’s not yet known is how this will affect model hallucination rates when schema loading is delayed or how caching efficiency compares across different deployment environments.

Context

Tool Retrieval could significantly lower the operational cost and complexity of deploying AI agents in enterprise and consumer applications, particularly where hundreds of tools are involved. By reducing token usage and latency, it may make AI-driven automation more viable for real-time systems like customer support or e-commerce workflows. Developers may adopt it to build more scalable agents without redesigning their tool architecture, potentially accelerating innovation in agent-based software. The feature may also influence how developers structure their tool libraries, favoring modular, on-demand access over monolithic integration. A comparable precedent is the introduction of vector databases in retrieval-augmented generation (RAG) systems, which similarly shifted from static to dynamic data access. However, unlike RAG, Tool Retrieval does not require developers to manage embeddings or indexes, lowering the barrier to entry. Still, questions remain about whether the added round-trip for client-side searches could offset gains in efficiency. The broader impact may be a shift toward more dynamic, adaptive AI agents—though widespread adoption will depend on how well the feature performs under load and whether it integrates smoothly with existing AI orchestration platforms.

Expanded detail and Context are AI-generated analysis; the linked article remains the authoritative source.
Read the full article at Google Blog →
Related stories
Gemini App Expands to Low-Memory Android Devices with Enhanced Features · Chatbots & Assistants
Apple Introduces Paid AI Usage Tiers for Pro Creative Apps · AI Coding Tools
Google Unveils Gemini Decisions API for Rapid Multimodal Decision-Making · AI Coding Tools
This summary is Al-enhanced to contain extended analysis and broader social context. The original is {NAME); the linked article is the authoritative source. Original headline: “Introducing Tool Retrieval in the Gemini API.” Browse more stories.