Google’s always-on AI agent can now book your dinner, restock your kitchen, and design your next presentation — all without you lifting a finger.
When Google unveiled Gemini Spark at I/O 2026 in May, the pitch was already ambitious: a personal AI agent that lives in Google’s cloud, runs around the clock, and keeps working even after you close your laptop. Now, as beta access rolls out to Google AI Ultra subscribers in the United States this week, the full scope of what Spark can actually do is coming into focus — and it extends well beyond Google’s own ecosystem.
Spark can now execute tasks in third-party applications. We’re not talking about reading a calendar invite or summarizing an email. We’re talking about booking a restaurant reservation through OpenTable, placing a grocery order on Instacart, or building a presentation template in Canva — autonomously, in the background, while you focus on something else entirely.
This is a meaningful leap, and it raises questions that go far beyond convenience.
What Gemini Spark Actually Is
At its core, Gemini Spark is a persistent cloud agent. It runs on dedicated virtual machines inside Google’s infrastructure, which means it doesn’t need your phone, your laptop, or even your attention to operate. While previous AI assistants were reactive — waiting for you to open an app and type a prompt — Spark is proactive. It monitors your Gmail, watches your calendar, tracks your ongoing projects in Google Docs, and surfaces actions you might want to take before you’ve thought to ask.
Users can interact with Spark by emailing a dedicated Gmail address or invoking it through the Gemini app. But the deeper value proposition is the background operation: Spark can be given a standing task — “make sure I have groceries for a dinner party this Friday” — and then figure out the steps, including placing an Instacart order, without any further input.
Built on Gemini base models with an agentic harness from Google Antigravity, Spark is designed to chain multi-step tasks across time. That’s fundamentally different from a chatbot that answers questions or a copilot that completes sentences.
MCP: The Key to Third-Party Access
The architecture powering Spark’s reach into outside apps is MCP, or Model Context Protocol — the open standard that has rapidly become the connective tissue of the agentic AI era. Spark currently supports more than 30 third-party integrations via MCP, with launch partners including Canva, OpenTable, and Instacart.
MCP allows Spark to interact with external services through a standardized interface, meaning developers can build Spark-compatible integrations without deep Google partnerships. Google has signaled it will continue expanding the list of supported apps, and developer tooling for MCP integration is already available.
This matters because it sets Spark up as a platform, not just a product. If the MCP ecosystem expands the way Google intends, Spark could eventually act on behalf of users across dozens of services — functioning less like a single app and more like an autonomous layer sitting on top of your entire digital life.
Paying for Things: The Agent Payments Protocol
Perhaps the most consequential new capability is Spark’s ability to complete purchases. Google is handling payments through what it calls the Agent Payments Protocol, or AP2 — a permission framework designed to give users control over what Spark can spend and on what.
Under AP2, Spark cannot complete a transaction without explicit user confirmation. The system supports per-transaction approval settings, meaning you can authorize Spark to place Instacart orders up to a certain dollar amount automatically, while requiring manual review for anything above that threshold.
It’s a sensible design for a first-generation payments system, though it also reveals something important about where Google is being careful. Autonomous AI agents that spend money create obvious risk — both financial and reputational. By building in a human confirmation layer by default, Google is essentially acknowledging that trust has to be earned incrementally. Whether users will actually engage with those confirmation prompts, or simply approve everything to reduce friction, remains to be seen.
Who Can Use It, and What It Costs
Spark’s beta is currently limited to Google AI Ultra subscribers in the United States. Ultra is Google’s top-tier AI subscription, recently reduced to $100 per month — a price cut that came alongside the Spark announcement and that many observers interpreted as Google trying to accelerate adoption ahead of competing agent products from OpenAI and Anthropic.
EU and UK users will have to wait longer. Compliance with the EU AI Act is still under review, and analysts expect Spark to reach those markets no earlier than Q3 2026.
The $100/month price point is steep for individual consumers, but squarely positioned for power users and professionals who already rely on Google Workspace. For small business owners, for example, an agent that handles scheduling, ordering, and document drafting in the background has an obvious value proposition — even at a four-figure annual price.
The Bigger Picture: Ambient AI Is Arriving
Gemini Spark’s expansion into third-party apps is a signal, not just a product feature. It represents Google’s opening move in a broader bet that the future of AI isn’t chatbots — it’s ambient agents.
The distinction matters. Chatbots are tools you use. Ambient agents are systems you live alongside. They operate on your behalf with persistence, context, and initiative. They know what you were working on last Tuesday and what you have coming up next Monday. They don’t forget context between sessions, because they never technically end a session.
This kind of always-on AI creates genuine value — it also creates genuine questions about data, privacy, and control. When an AI agent has standing access to your email, calendar, documents, and payment methods, the scope of potential exposure is significant. Google has so far been relatively forthcoming about Spark’s privacy architecture, noting that agent activity is stored in a user-controlled log and can be audited or revoked. But the details of data retention, training use, and third-party data sharing will require sustained scrutiny as the product matures.
What Comes Next
Google has indicated that additional third-party integrations are coming throughout the summer, with a particular focus on productivity tools, travel services, and e-commerce platforms. Support for voice-initiated tasks through Android is expected later in 2026, which would significantly expand the surface area for Spark interactions.
Competitors are not standing still. OpenAI’s Operator product, which also enables browser-based task execution, has been expanding its capabilities steadily. Anthropic has signaled its own agentic product roadmap. The race to build the default ambient AI layer is well underway.
For now, Gemini Spark represents the most technically polished version of the ambient agent idea that’s been deployed at scale. Whether it earns the trust — and the habits — of its early users will determine whether Google’s cloud-first, always-on approach becomes the template for how AI agents work, or a costly lesson in how much farther this technology still has to go.
Beta access for Gemini Spark is available to Google AI Ultra subscribers in the US at $100/month. Third-party integrations currently include Canva, OpenTable, and Instacart, with more announced for summer 2026.
