Google‑Style Tool Orchestration

Most application programming interface tool calling still works like a relay race: the model calls Search, your code parses the result, then your code starts a second request for Maps or a custom function. Google’s March 17, 2026 Gemini update lets developers put built-in tools and custom functions into one `generate_content` call instead. (blog.google) A built-in tool is a capability the model already knows how to use, like Google Search or Google Maps. A custom function is your own code, like `getWeather` or `bookTable`, that the model can call by name with structured arguments. (ai.google.dev, ai.google.dev) The hard part is not calling one tool once. The hard part is keeping the whole trail of what happened so the model remembers which search result led to which map lookup and which function response belongs to which request. (ai.google.dev) Google’s answer is “context circulation,” which means the history of tool calls, tool responses, and model state gets passed back into the next turn instead of being flattened into plain text. The Gemini documentation says official software development kits handle this automatically when you use standard chat history. (ai.google.dev, developers.googleblog.com) That matters because newer Gemini models also return “thought signatures,” which Google describes as encrypted representations of the model’s internal reasoning. Developers do not read those signatures, but passing them back helps the model continue the same chain of reasoning across multiple steps. (developers.googleblog.com) Google also added explicit IDs for tool calls and tool responses, including parallel cases where the model asks for more than one function at once. Those IDs work like baggage tags at an airport: they tell the system which response belongs to which request when several are moving at the same time. (ai.google.dev) The new piece in this story is Google Maps grounding. Google’s current Gemini tool docs list Maps alongside Search, Uniform Resource Locator context, code execution, and computer use as built-in tools that can feed live external data into a model response. (ai.google.dev) In practice, that lets one request do a chain like this: search for a place, pull map details for that place, then call a company function that checks inventory, pricing, or reservations. MarkTechPost’s April 7, 2026 walkthrough shows that pattern in five demos that move from a simple search-plus-function example to a fuller multi-tool chain. (marktechpost.com) This does not mean one request magically removes orchestration code. The Gemini docs still show developers defining function schemas, executing the returned function calls in their own code, and then sending the function responses back so the model can continue. (ai.google.dev, ai.google.dev) What changed is where the complexity sits. Instead of every team inventing its own glue for search results, map lookups, response matching, and multi-step memory, more of that behavior is now part of the vendor’s tool semantics and official software development kits. (blog.google, developers.googleblog.com) That is why this looks bigger than a single feature launch. When model vendors expose richer tool behavior inside one call, the winning internal platform teams will be the ones that turn those raw capabilities into safe, reusable wrappers so product teams can ship agents without rebuilding the plumbing every time. (ai.google.dev, blog.google)