[I/O 2026] Gemini Spark, the personal agent that continuously performs tasks

Google unveils Gemini Spark, an autonomous agent powered by Gemini 3.5 Flash and Antigravity that executes background tasks via Google and third-party apps.

Google unveils Gemini Spark, an agent designed to act autonomously on the user's digital life while remaining under their control. It represents an evolution of Gemini: from a reactive assistant to a system capable of executing concrete tasks for extended periods.

Spark leverages Gemini 3.5 Flash and the Antigravity environment. Hosted in the Google cloud, it continues to operate in the background on virtual machines, even when the user's computer is turned off or their phone is locked.

It can connect to common Google applications (Gmail, Calendar, Drive, Docs, Sheets, Slides). These accesses are disabled by default and must be explicitly authorized by the user. Its operation relies on three main elements: delegated tasks (specific actions entrusted by the user); skills (recurring automations defined by the user); triggers (scheduled or conditional).

Spark also integrates new connections via the MCP (Multi-Connector Protocol), with an initial group of third-party applications: Canva, OpenTable, and Instacart. Further integrations are underway.

Among the features announced for the coming months are command by message or email, the creation of personalized sub-agents, and local browser control.

Before any sensitive action (payment, email sending, significant document modification, etc.), Spark requests explicit user validation.

The rollout begins with a restricted group of trusted testers. It will be followed by a beta reserved for Google AI Ultra subscribers (of legal age) in the United States.