As agentic workflows scale, orchestration can develop into a bottleneck. At the moment, we’re simplifying this by permitting builders to combine built-in instruments (akin to Google Search and Google Maps) with customized features in a single request, circulate context throughout software calls and turns for extra advanced reasoning and lengthen Grounding with Google Maps to the Gemini 3 mannequin household.
New tooling capabilities within the API
Mix built-in and customized instruments in the identical interplay
Beforehand, builders needed to fastidiously orchestrate when to make use of built-in instruments (like Google Search) versus when to depend on a customized operate declaration. Now, you possibly can pass both built-in tools and your own custom tools in the identical request. This enables Gemini to simply pivot between fetching public information through Google Search then calling your backend with out separate orchestration steps, decreasing end-to-end latency and simplifying agent architectures.
This has been a high request from builders since we launched built-in instruments and we’re excited to see the way you mix file search, Google Maps, Search, and customized features collectively!
Cross-tool context circulation for built-in instruments
In multi-step workflows, fashions usually want to make use of the output of 1 software because the enter for an additional. Context circulation for built-in instruments preserves each software name and its response within the mannequin’s context, so follow-up steps can entry and cause over that information. For instance, Gemini can now use a built-in software to get real-time climate information and flow into that context to a customized software that books a venue.
Device response IDs
To enhance debuggability and guarantee exact mapping throughout asynchronous software executions, we’ve launched unique call identifiers (`id`) for each software name. These IDs enable builders to establish particular software calls requested by the mannequin with the precise consumer responses, which is very essential when dealing with parallel operate calling and cross-tool context.
Right here’s a code snippet displaying an instance of a multi-tool mixture stream with Grounding with Google Search.
