Google formalises Gemini stack

Google is turning Gemini from one chatbot into a whole tool belt. In the last week, signs of a wider rollout showed Google pushing “Skills” into more parts of Gemini, adding specialist subagents for developers, and folding NotebookLM’s research workflow into the main Gemini app. (testingcatalog.com, geminicli.com, techrepublic.com) A Skill is basically a saved playbook. Instead of retyping the same long instruction every time, a user can keep a reusable set of rules and call it when needed, and TestingCatalog says Google is preparing to expand that idea into Google Artificial Intelligence Studio as well as Gemini itself. (testingcatalog.com) A subagent is a different piece of that stack. Google’s Gemini command line interface docs describe subagents as specialist agents that work inside a main session, so one agent can handle a narrow job like documentation lookup or codebase analysis without stuffing every detail into the main conversation. (geminicli.com, docs.cloud.google.com) That split tells you what Google is building. Skills are reusable instructions, while subagents are reusable workers, which means Gemini is being shaped less like a single assistant and more like a manager that can hand jobs to the right specialist. (testingcatalog.com, geminicli.com) Google is making the consumer side match that structure too. This week it started bringing NotebookLM-style notebooks into Gemini, so chats, files, notes, and instructions can sit in one project hub instead of being split across separate apps. (techrepublic.com, digitaltrends.com) Google has been widening the language layer around that hub for a while. In February 2025, Gemini in the side panel of Google Docs, Google Sheets, Google Drive, and Gmail expanded to 17 more languages, including Arabic, Chinese, Polish, Thai, Turkish, Ukrainian, and Vietnamese. (workspaceupdates.googleblog.com) Now Google is testing whether that same system can do errands, not just writing. In India, Google Search’s Artificial Intelligence Mode has started helping users find and book restaurant tables by taking one request with details like party size, cuisine, date, and time, then checking live availability across Zomato, Swiggy, and EazyDiner. (business-standard.com, gadgets360.com) That is the same architecture showing up in public. A user gives one broad instruction, Gemini breaks it into smaller tasks like search, comparison, and booking, and Google keeps the whole flow inside its own interface instead of sending people app to app. (business-standard.com, testingcatalog.com) The awkward part is that a more modular Gemini also creates more places to leak money or data if developers wire it up badly. CloudSEK said this week that it found 32 hardcoded Google application programming interface keys in 22 Android apps with more than 500 million installs combined, and those keys could expose access to Gemini endpoints, uploaded datasets, cached contents, and billable usage. (cloudsek.com, itnerd.blog) So the story is not just that Google added another assistant feature. Google is formalising a full Gemini stack: saved playbooks for repeat jobs, specialist agents for narrow tasks, research spaces inside the main app, and real-world actions in markets like India, all while the security plumbing underneath is suddenly much more important. (testingcatalog.com, geminicli.com, techrepublic.com, cloudsek.com)