Google’s Gemini Spark: What Its New 24/7 Personal AI Agent Actually Does

At Google I/O 2026 on May 19, Google announced Gemini Spark, a personal AI agent that works differently from the chatbots most people are used to. Instead of waiting for a prompt and replying once, Spark runs continuously in the background — Sundar Pichai described it as “your personal AI agent that helps you navigate your digital life, taking action on your behalf and under your direction.” It keeps working even after you close the app, lock your phone, or shut your laptop, checking your inbox, tracking deadlines, and handling multi-step tasks on its own until it needs your input.

That “always-on, takes action” framing is the real shift. This post covers what Spark actually does, who can use it and where, how it stacks up against Anthropic’s Claude Cowork and OpenAI’s ChatGPT Agent, and — because our position on this site is to skip the hype and give you the straight version — a warning Google itself has issued about it that’s worth taking seriously before you turn it on.

What Gemini Spark Actually Does

Spark is built on Gemini 3.5 and runs on Antigravity, Google’s agentic infrastructure platform. Rather than living inside a single chat window, it runs on dedicated Google Cloud virtual machines, which is what lets it keep working after your device is off — the task isn’t paused on your phone, it’s running on Google’s servers.

At launch, Spark connects to a specific set of Google apps and services through direct API integrations rather than screen-scraping: Gmail, Google Calendar, Drive, Docs, Sheets, YouTube, and Google Maps. In practice, that means it can:

• Draft email replies and monitor your inbox for things that need attention
• Track deadlines across your Calendar and Docs and remind or act on them
• Pull together information spread across Drive, Sheets, and Docs into one task
• Work on a task in the background while your phone is locked or your laptop is closed

Google also lists three third-party services connected via MCP (Model Context Protocol) at launch — Canva, OpenTable, and Instacart — so Spark’s reach extends a little beyond Google’s own apps, though the core value proposition is squarely about what it can already see and touch inside your Google account.

Availability and Pricing

Spark rolled out cautiously. Google’s first release went to trusted testers, followed by a beta for Google AI Ultra subscribers — the $100/month plan not-so-coincidentally introduced at the same I/O 2026 keynote, alongside 5x higher usage limits than Pro, YouTube Premium, and 20TB of cloud storage.

To use Spark at all, Google requires a personal Google Account (work and school accounts aren’t currently supported), that you be 18 or older, and that Keep Activity is turned on. Geographic availability is broad but not universal — as of this writing it’s unavailable in the European Economic Area, the United Kingdom, Switzerland, and Nigeria.

Access has since expanded beyond Ultra: Google AI Pro subscribers can now get Spark too, though currently only in the United States and only in English, while Ultra subscribers get broader language and country coverage (subject to the exclusions above). India got its own rollout on July 29, 2026 — Ultra subscribers there already had access via the Gemini app and website, and Google began phasing Spark out to Pro subscribers in India the same week.

How It Compares to Claude Cowork and ChatGPT Agent

Spark isn’t the only persistent AI agent that launched around this time. Anthropic has Claude Cowork, currently a research preview limited to Max subscribers on Anthropic’s $100–$200/month plans, and OpenAI has ChatGPT Agent, bundled into existing Plus, Pro, Business, and Enterprise plans at no extra list price.

The three take genuinely different approaches rather than competing on the same axis:

• Gemini Spark’s edge is depth of integration with data Google already has access to. Because it connects to Gmail, Calendar, Docs, and Drive through direct APIs rather than reading your screen, it’s well-suited to background inbox management and calendar-driven tasks for anyone who already lives inside Google Workspace — at the cost of being the priciest entry point of the three, and being Google-ecosystem-bound.
• Claude Cowork is generally regarded as stronger for producing finished files — spreadsheets, slide decks, documents, PDFs — and research deliverables, which suits work-product-focused tasks rather than ongoing inbox monitoring.
• ChatGPT Agent is the lowest-friction way in, since it’s included in plans people are often already paying for, and leans toward browser-native tasks.

None of this makes one strictly “better” — the honest takeaway is that they’re built for different jobs, and picking one over another mostly comes down to which apps your work already lives in.

An Important Caveat Google Itself Has Raised

This is the part we’re not going to soften. Spark’s own onboarding screen, reported by multiple outlets including Forbes, tells users directly:

“While it is designed to ask for your permission before taking sensitive actions, it may do things like share your info or make purchases without asking. Make sure to supervise Gemini Spark, and don’t rely on it for medical advice, legal, financial, or other professional help.”

In other words, Google’s own documentation acknowledges that an agent explicitly “designed to ask for your permission” before sensitive actions can still skip that step and share your information or spend your money without confirmation. Reporting also noted Google softened this onboarding language somewhat before the wider launch, but the core warning survived into the released product — this isn’t a leaked early draft, it’s what ships.

Spark draws context from your connected apps, past chats, tasks, signed-in websites, and location to decide what to do, and it can route some of that data to third parties to complete a task. Google’s stated mitigations are real but manual on your part: you choose which apps it can access, you can disable Connected Apps, you can clear its activity, and you can turn it off entirely. None of that is automatic — it depends on you actively supervising it rather than setting it up once and forgetting about it, which is a meaningfully different trust model than a chatbot that only acts when you type something.

An Illustrative Example

The scenario below is illustrative only — a composite example built to show how Spark’s pieces fit together in practice, not a real user, case study, or verified outcome.

Picture a freelance consultant juggling several client threads in Gmail. They turn on Spark and grant it access to Gmail, Calendar, and Drive. Over the course of a week, Spark flags a client email that’s been sitting unanswered for three days and drafts a reply for review rather than sending it outright. It notices two meeting requests that overlap on the calendar and surfaces the conflict instead of picking one itself. When a client asks for “the proposal doc,” Spark locates the right file in Drive and attaches it to a drafted reply. In each case, the consultant reviews and approves before anything goes out — which is exactly the kind of active supervision Google’s own onboarding warning says is necessary, not optional.

Frequently Asked Questions

Want to explore more on how AI tools are changing everyday work? Browse the rest of our AI category, or try one of our free AI Tools.