Google's New AI Keeps Working Even After You Walk Away

Google has unveiled a new kind of artificial intelligence that doesn't stop when you do. At its I/O 2026 event, the company introduced Gemini Spark, a persistent digital assistant designed to handle multi-step tasks across email, documents, calendars and the web. Unlike standard chatbots that reply once and go silent, Spark runs in the background on Google Cloud servers. You can close your laptop or lock your phone. The agent keeps going.

Alphabet CEO Sundar Pichai described it as a personal AI that helps navigate your digital life, taking action under your direction. The system runs on Gemini 3.5 Flash and uses technology from Google's Antigravity project. Josh Woodward, a Google vice president, said Spark feels like tossing tasks over your shoulder and having them caught and completed.

Early testing began this month with a small group. A broader beta is now open to Google AI Ultra subscribers in the U.S., who pay roughly $100 per month, plus select business users. The agent connects to Gmail, Calendar, Drive, Docs, Sheets and Slides, but those connections are turned off by default. Users must activate them. The agent does not scan inboxes randomly. It acts only on assigned tasks.

Practical uses include scanning email threads for action items and building a prioritized to-do list. It can log receipts from group-trip planning into a spreadsheet and email the updated plan. Freelancers can ask it to extract client details from inquiries, create a tracker row and set up a dedicated Drive folder. Homeowners get reminders about upcoming maintenance based on past invoices. The agent can also browse multiple websites, compare options and help with bookings.

Third-party connections will expand through the Model Context Protocol, an open standard. Planned partners include Canva, OpenTable, Instacart, Spotify, Expedia and Adobe. Later this year, Spark will access local files on Mac computers. Users will eventually text or email the agent directly.

WIRED writer Reece Rogers put the beta through a real-world test. He gave Spark access to his Gmail, Docs and Calendar, then asked it in one vague sentence to plan his birthday party. The agent delivered a five-page itinerary within minutes. It found an actual karaoke reservation, compiled a guest list capped at the room's capacity, outlined venue rules, suggested dinner options and drafted follow-up emails. The list ranked Rogers' live-in boyfriend as a close friend but omitted Rogers himself from his own celebration. Spark chose specific queer bars based on keyword matches in past receipts rather than any understanding of identity. When Rogers tested email drafting, the agent mimicked his tone but required manual fixes for reservations that glitched. Rogers ended up calling the sushi restaurant himself.

Google stresses user oversight. The agent checks before major actions like sending emails or initiating payments. Users can set repeating schedules, upload custom skills that learn email style from past messages, or interrupt at any time. Security warnings appear prominently. A malicious prompt could instruct the agent to extract private data and post it publicly. The company advises caution with full inbox access.

Analysts watch closely. If Spark delivers consistent value on routine work, adoption could accelerate. Persistent glitches or privacy incidents would slow it. The architecture marks a shift from on-device assistants to cloud-resident workers that operate independently yet remain accountable. Executives describe an active partner rather than an autonomous replacement. The system still needs clear direction. It excels at synthesis and execution but struggles with implicit social cues.

Privacy remains the largest open question. Deep access to email, calendars, documents and soon local files hands Google an unprecedented view of daily routines. The company says connections require explicit opt-in and that the agent does not read mail constantly. Yet the always-on design invites scrutiny. Early feedback suggests the experience feels magical when it works and jarring when it doesn't.

One thing looks certain. The era of agents that simply answer questions has ended. The agents that act, persist and occasionally surprise their owners have begun.