Google is internally testing a 24/7 AI agent called Gemini Spark in the web version of Gemini. Developer @mrfanduuuuu today on X posted screenshots of its onboarding page, marked as BETA, expected to be officially unveiled at the I/O keynote on May 19.
Spark’s core selling point is that 'the more you use it, the more it understands you.' It pulls context from connected Google apps, skill modules, chat history, scheduled tasks, logged-in websites, Personal Intelligence (Google’s personal data integration feature launched in January), and location data. Then it handles multi-step tasks like inbox management, web operations, form filling, and shopping checkouts. During execution, Gemini shares your name, contact info, files, and preferences with third parties.
But the risk warning on the onboarding page is blunt: Spark is still experimental. While it’s designed to ask for permission before sensitive actions, it might share your info or complete a purchase without your confirmation. To maintain session continuity, the system also saves remote browser data, including login credentials and remote code execution data. You can clear this data in settings and turn off Connected Apps.
According to TestingCatalog, Spark is the evolution of the internal project codenamed 'Remy' — an upgraded Gemini Agent previously only available to AI Ultra subscribers. The same leak also revealed a new model called 'Spark Robin' (described as 'Rich Visual Response') and a new 'MCP Tool Testing' category in the Gemini model selector, hinting that Spark will connect to third-party tools.
With this always-on agent, Google is directly challenging Anthropic’s upcoming hosted agent Conway and OpenAI’s agent platform. The competition among the three has shifted from 'whose model is smarter' to 'who can do more work for you.'