Google Deepmind upgrades Gemini API with multi-tool chaining and context circulation - The Decoder
Google is trying to solve one of the most persistent problems in agentic AI systems: the memory gap between tool calls.
When a multi-step AI agent calls a tool — Google Search, a custom database, a Maps lookup — each call happens in isolation. The model gets the result, but the trail of what it did and why gets lost in long conversations. The system that called Google Search in step three doesn't automatically share that result with the system calling the custom booking function in step four. Developers spend significant engineering time building middleware to stitch these steps together.
Google's answer, announced Tuesday on the Google Developers Blog, is called context circulation. The new feature preserves every tool call and its response in the model's context window, so follow-up steps can access and reason over that data directly. The example Google gives: Gemini uses a built-in weather tool, then circulates that context to a custom tool that books a venue — without the developer having to manually wire the output of one call into the input of the next.
The update is part of a broader set of API changes aimed at reducing orchestration complexity in agentic workflows. The other major addition: developers can now pass both built-in tools (Google Search, Google Maps) and their own custom functions in a single API request, where previously these required separate orchestration steps. Google is also extending Grounding with Google Maps to the Gemini 3 model family, giving models access to spatial data, local business information, commute times, and place details as part of their reasoning.
The company is also recommending developers use a new Interactions API for these workflows, which handles server-side state management and provides unified reasoning traces — a signal that the current generateContent API was not designed for the long, multi-step tool chains that production agents require.
Tool response IDs are another new addition: every tool call now gets a unique identifier that maps it precisely to its response, solving a debuggability problem in asynchronous and parallel function calling.
The practical effect is that building complex agentic workflows requires less custom infrastructure. The model does more of the orchestration work directly. For developers building on Google's platform, this is a meaningful reduction in the gap between "the model can do X" and "the system can reliably do X in production."
Sources: