Google Gemini Spark launches as 24/7 AI agent for AI Ultra subscribers

Google announced Gemini Spark on May 19 at I/O 2026 as a cloud‑resident AI agent that will be available next week to U.S. AI Ultra subscribers for $100 a month [2]. Spark runs on the new Gemini 3.5 Flash model inside Google’s Antigravity 2.0 runtime, letting the agent execute multi‑day tasks, act on user commands and stay active even when devices are offline [2].

The code analysis of Android’s Google App version 17.20 shows that the on‑device Gemini Intelligence feature already includes a background service, a skill system and a task scheduler [1]. Strings such as assistant_robin_agent_tab_schedules and assistant_robin_agent_tab_skills point to planned capabilities for scheduled, modular actions. The service can run continuously in the foreground (ROBIN_AGENT_MODE_FG_SERVICE_STARTED) and is gated by a multi‑layer security check that currently limits full functionality to Pixel 10 series and Galaxy S26 devices with Android 16 or higher [1].

Spark extends these on‑device ideas to the cloud. By using the Model Context Protocol (MCP), Spark connects to more than 30 third‑party services—Canva, OpenTable, Instacart, and Google Workspace apps—without exposing raw credentials to the language model [2]. A desktop client indexes authorized local folders and feeds them to the cloud agent via the same MCP sandbox, enabling actions like drafting emails while the user sleeps. Google’s rollout is limited to the United States, with EU availability delayed pending regulatory clarity under the AI Act [2].

The shift from a device‑bound assistant to a persistent cloud agent marks a strategic move: Google can now offer a subscription‑based AI service that performs proactive tasks, such as monitoring flight prices or managing inboxes, without requiring the user’s device to be active. Whether the $100/month price point will attract enough power users to justify the infrastructure remains the key question as Google expands Spark’s skill ecosystem.