Google shifts to agentic AI workflows

Google used its May 19 I/O announcements to recast Gemini as a system for doing work, not just answering prompts. The company introduced Gemini 3.5, launched a proactive assistant called Gemini Spark, and added managed agents to the Gemini API for developers. Across those releases, Google described the same goal in different ways: software that can carry out multi-step tasks across apps, files and context under user direction. The shift matters because Google is tying its consumer app, developer tools and Workspace products to the same “agentic” model. In Google’s wording, Gemini 3.5 is a family of models that combines “frontier intelligence with action,” while Spark is a “24/7 personal AI agent” and the Gemini API now supports managed agents in a secure cloud sandbox. ### So what actually changed in Gemini itself? (blog.google) Google said on May 19 that Gemini 3.5 is its latest model family and that it is starting the series with Gemini 3.5 Flash. The company said 3.5 Flash is designed for speed as well as tool use, and framed it as the engine for “real-world agentic workflows.” TechCrunch reported that Google made Gemini 3.5 Flash the default model in the Gemini app and in AI Mode in Search globally, underscoring that the company is pushing agents rather than standard chatbots as its next product wave. (blog.google) ### What is Gemini Spark supposed to do that chatbots do not? Google described Gemini Spark as a 24/7 personal AI agent that can proactively manage tasks, with the user still directing what it does. (blog.google) In the Gemini app update, Google paired Spark with a new “Daily Brief” agent and said the app is becoming more agentic and more proactive. Google also connected Spark to productivity products. (techcrunch.com) In its Workspace announcement, the company said Spark would sit alongside updates in Gmail, Docs and Keep, linking the agent push to everyday office software rather than treating it as a standalone demo. ### What did Google give developers beyond a new model? Google said the Gemini API now supports managed agents that run in a secure cloud sandbox. (blog.google) Developers can define those agents with versionable files including `AGENTS.md` and `SKILL.md`, then attach instructions, skills and data to automate workflows. That structure is notable because Google is packaging agents more like software projects than prompt templates. (blog.google) The company’s developer posts also tied the release to Antigravity, its agent-first development platform, and said the broader aim is to move “from prompts to action.” ### Where is Google aiming these agents first? Google’s own examples point to document-heavy knowledge work. (blog.google) The company’s app and Workspace posts focus on organizing information, managing tasks and helping users get things done across Gmail, Docs and related tools. TechCrunch’s reporting pointed to the same use case in plainer terms: workflow automation instead of one-off answers. (blog.google) In practice, that means tasks such as assembling reports, extracting action items from meetings, or moving information across tools with less manual copying and prompting. That is an inference from Google’s product descriptions and TechCrunch’s reporting on the launch direction. (blog.google) ### How should readers think about this push? Google’s May 19 releases show a coordinated product move across models, consumer software and developer infrastructure. The company did not present agents as a separate niche category; it embedded them in Search, the Gemini app, Workspace and the Gemini API on the same day. The next concrete markers are already public. (blog.google) Google’s May 19 blog posts for Gemini 3.5, the Gemini app, Workspace updates and managed agents lay out the features now being rolled out, and TechCrunch reported that Gemini 3.5 Flash is already becoming the default in key Google surfaces. (blog.google)