Gemini Spark launches with dedicated VMs and MCP support for 24/7 background agents

• Google introduced Gemini Spark as a persistent agent inside Gemini that can keep working after you close your laptop, according to Google's launch post and Google's product demo post.
• The core infrastructure claim is unusually concrete: Spark runs on dedicated virtual machines on Google Cloud, while testingcatalog's keynote screenshot says the stack is Gemini 3.5 plus Google's Antigravity harness.
• Google is framing Spark as a tool-connected agent from day one, with Google product integrations now and third-party tool access through MCP soon, per Google's launch post and GeminiApp's roadmap post.
• Access is staggered. GeminiApp's rollout thread says trusted testers get it this week, U.S. Google AI Ultra subscribers get a beta next week, and a business version is planned for Workspace and Gemini Enterprise.
• The roadmap is already broader than the launch pitch: GeminiApp's summer roadmap lists MCP integrations, macOS availability, text and email access, Chrome browser control, and subagents.

A day before launch,

described Spark as an always-on assistant that learns from connected apps and skills. By keynote time, testingcatalog's stage photo confirmed the dedicated VM, Antigravity, and MCP pieces, and GeminiApp's follow-up thread added a desktop angle, including local file organization and PDF-to-Sheets extraction.

Google is pitching Spark as the shift from Gemini as a chat assistant to Gemini as a delegated worker. The official description is a 24/7 personal agent that takes actions on your behalf, under your direction.

The launch materials make three product claims:

• Runs continuously in the background
• Connects to Google tools now
• Expands to third-party tools through MCP soon

The most important implementation detail is not the agent label, it is the execution model. Google says Spark runs on dedicated virtual machines on Google Cloud, which is why Google's closed-laptop post can market the whole thing with "Laptop? Closed. Spark? Running."

That architecture puts Spark closer to a hosted agent session than a foreground assistant tab. testingcatalog's keynote screenshot also says Google built it on Gemini 3.5 and the Antigravity harness, tying Spark to the broader agent stack Google introduced at I/O.

The rollout starts small. GeminiApp says trusted testers get Spark this week, while U.S.-based Google AI Ultra subscribers get a beta next week. The same thread says Google is also planning a business variant for Workspace and Gemini Enterprise.

Google also widened the surface area after the keynote. In GeminiApp's desktop follow-up, the company says Spark is coming to the Gemini desktop app for macOS, where it will handle tasks like organizing local files and extracting PDF data directly into Google Sheets.

That desktop detail lines up with the prelaunch leak, where testingcatalog's reply about local folders said Spark would work on attached local folders.

Google showed a five-item roadmap for the summer:

1. MCP integrations
2. macOS availability
3. Text or email Gemini Spark
4. Google Chrome browser control
5. Create subagents

The subagents item is the standout. Google launched Spark as a single personal agent, but the roadmap already points toward multi-agent task decomposition inside the consumer Gemini app.