Gemini Spark

Gemini Spark is Google's answer to the question: what happens when an AI agent never stops working? Announced at Google I/O in May 2026 and now rolling out to AI Ultra subscribers in the US, Spark runs continuously on Google Cloud virtual machines. You assign it a task, close your laptop, and it keeps going. It can scan your inbox, pull context from Docs and Sheets, browse the web through Chrome, and complete multi-step work without you being present for each step.

What sets Spark apart from earlier chat-based assistants is its integration depth. Because Google controls Gmail, Calendar, Drive, and Chrome, Spark can move between those surfaces without the friction that standalone agents face when stitching together third-party tools. You can even email Spark directly at a dedicated Gmail address, the way you'd message a colleague. It also supports MCP (the open standard for connecting agents to external services), so it reaches beyond Google's own apps too.

Spark is built around three primitives: Tasks (what you want done), Schedules (when or under what conditions to do it), and Skills (reusable instructions that shape how Spark behaves over time). The model powering it is Gemini 3.5 Flash running inside Google's Antigravity agent harness. Before taking high-stakes actions like sending an email on your behalf, it's designed to check in with you first. Privacy trade-offs are real: Spark requires opt-in access to your personal data across Google services, and that's a decision worth making deliberately.