AI assistants have been stuck in a Q&A loop — you ask, they answer. Google wants to change that with Gemini Spark, a proactive agent announced at I/O 2026. It runs on the Gemini 3.5 model, uses a new Antigravity framework, and taps into Workspace. But the big deal: Spark can keep working in the cloud even after you shut your computer or lock your phone.
Unlike traditional passive assistants, Spark is built for "proactive tasking" and "offline cloud execution." You can teach it skills or set up workflows: for example, trigger it to scan monthly bills for hidden subscription fees, monitor your inbox for deadlines and compile a daily digest for your family, or stitch together meeting notes from emails and chats into a Google Doc with a launch email ready to send.
Today, Google also introduced MCP (Model Context Protocol) connectors for Spark, an open standard that lets the model read external data and call tools. The first integrations are with Canva, OpenTable, and Instacart — within weeks, Spark will be able to place orders or create designs across these apps on your behalf.
Google is drawing a hard safety line: any high-risk action like paying money or sending emails requires explicit user confirmation.
Gemini Spark starts trusted tester access this week, with a beta for Google AI Ultra subscribers in the US arriving next week. This summer, it will also land on the new macOS desktop client to take over local file and system workflows.