Gemini Spark: Google's 24/7 AI Agent Explained

Gemini Spark is Google's biggest bet on artificial intelligence this year: an agent that works for you 24 hours a day, even with your laptop closed. Announced at Google I/O 2026, it doesn't answer questions — it executes tasks. I've gathered here what was confirmed at the event, what changes in practice, and what still warrants caution before you delegate your digital life to a robot.

TL;DR — what you need to know

• Gemini Spark is a personal 24/7 AI agent announced at Google I/O 2026 on May 19.
• It runs on Google's cloud, on dedicated virtual machines — works even with the device turned off.
• It is powered by the Gemini 3.5 model and the agentic harness inherited from Google Antigravity.
• It performs real tasks: reads credit card invoices, summarizes your inbox, and assembles documents in Google Docs.
• It launches first as a beta for Google AI Ultra subscribers in the US, on plans of $100 and $200 per month.

What is Gemini Spark

To understand Gemini Spark, it's worth separating two things that are often confused. The Gemini you already know — the chatbot within the app or Workspace — responds when you ask. Gemini Spark does the opposite: it acts without you watching.

On stage at Google I/O 2026, CEO Sundar Pichai defined the product, according to Google's official announcement, as "your personal AI agent that helps navigate your digital life, acting on your behalf." The key word there is agent. An assistant waits for instruction. An agent receives a goal and decides the steps.

In practice, Gemini Spark is an AI agent that works 24 hours a day, 7 days a week. You describe what you want — "organize the travel confirmations," "let me know if a new charge appears on my card" — and it executes in the background. When it finishes, it returns the ready result. It's a category change, not a version: Gemini has gone from a tool you operate to a collaborator that receives tasks.

How Gemini Spark works

What makes Gemini Spark different from ordinary automation is where it runs. Instead of depending on your device, each agent receives a dedicated virtual machine on Google Cloud infrastructure. This has an important practical consequence: you can close your laptop, lock your phone, or simply sleep — Gemini Spark keeps working.

Under the hood, there are two components. The first is the model: Gemini Spark uses Gemini 3.5, the generation of models Google presented at the same I/O 2026. The second is what Google calls the agentic harness — the layer that transforms a language model into something capable of planning, executing steps, and correcting its own course. This harness comes from Google Antigravity, the company's agentic development platform.

The difference between "model" and "harness" is not just technical detail. The model understands and generates text; the harness gives it hands. It's the combination of the two that allows Gemini Spark to open Gmail, read a message, decide it requires a response, and draft a reply — a sequence of decisions, not a single response.

What Gemini Spark already does today

Agentic features often sound abstract. That's why it's worth looking exactly at what Google demonstrated Gemini Spark can do today, without future promises.

Hidden subscription hunting

Gemini Spark can read your monthly credit card statement and flag new charges — especially those subscriptions you forgot you had. For anyone who has paid months for an unused service, this single feature already justifies looking at the product with attention.

Inbox summary

Instead of you opening dozens of emails, Gemini Spark monitors your inbox, identifies what changed, and sends a consolidated summary. Google demonstrated the case of school communications: the agent gathers teacher notices, extracts deadlines, and delivers a single digest with what matters. The same applies to a small business that needs to track orders without being glued to email.

Ready documents from conversations

The third feature is the most ambitious. Gemini Spark synthesizes information scattered across emails and chats and assembles an organized Google Doc — a meeting minutes, a project summary. It cross-references different sources and delivers the document already formatted, not a raw draft.

Gmail, Chrome, and MCP: the integrations that matter

An agent is only useful if it can reach your tools. That's where Gemini Spark differentiates itself from previous agentic AI experiments.

The most talked-about integration is Gmail. Gemini Spark gets its own email address: you write to your agent as you would to a human assistant. You ask for a task by email, it responds by email. Add to that native access to Google Docs, Sheets, and Slides, and the agent operates within the Workspace that many companies already use.

Chrome serves as the gateway to the rest of the web: Gemini Spark can browse pages directly through the browser, freeing it from the limits of closed integrations. On Android, the new Halo system shows the agent's progress in real time — you follow what it does without needing to interrupt.

Finally, Google opened Gemini Spark to MCP (Model Context Protocol), the open standard that connects agents to external services. At launch, connections with Canva, OpenTable, and Instacart were included. MCP is the same protocol that business platforms have been adopting to expose their own functions to agents — and it's through it that Gemini Spark should gain dozens of integrations without Google having to build each one by hand.

Price, plans, and availability

Here comes the reality check. Gemini Spark doesn't arrive for free nor for everyone at the same time.

The launch is phased. In the week of the announcement — May 19, 2026 — Gemini Spark was released to a restricted group of Google's trusted testers. Starting the following week, it enters beta for Google AI Ultra subscribers in the United States. The desktop app for macOS, capable of manipulating local files and automating workflows on the computer, is scheduled for the American summer of 2026.

Access is tied to the Google AI Ultra plan, which Google restructured at the same event. Here's how the plans look:

Plan	Price per month	Includes Gemini Spark	Notes
Google AI Pro	More affordable	No	Access to Gemini, without the agent
Google AI Ultra (entry)	$100	Yes (beta, US)	20 TB storage; limits 5× larger than Pro
Google AI Ultra (top)	$200	Yes (beta, US)	Was $250; expanded limits and early access

Two points deserve attention for those in Brazil. First: the beta is exclusive to the US for now — no official date for other countries. Second: even the entry plan costs $100 per month, which, converted to reais, is not trivial. Gemini Spark, at this moment, is clearly a product for early technology adopters.

The roadmap through the end of 2026

Google detailed what's still coming for Gemini Spark throughout the second half of the year. Planned features include sending text messages or emails directly to the agent, creating custom sub-agents — agents specialized in specific tasks — and authorizing payments by setting a budget and allowed stores. It's this last function, spending real money, that will measure how much trust the public is willing to place in an agent.

What Gemini Spark changes for Brazilian companies

I've worked with web infrastructure and automation for over 15 years, and few launches have seemed as game-changing as this one. Gemini Spark points to where corporate AI is heading: from answers to execution.

For a Brazilian company, the immediate impact is not using Gemini Spark tomorrow — it hasn't even arrived in the country. The impact is the pattern it sets. Customer service, email triage, sales follow-up: repetitive tasks that today consume people will increasingly be delegated to agents. I've already written about how the model of charging per customer service agent became obsolete — and the agentic logic accelerates this change: when the "agent" is software, the cost of adding another one plummets.

There's also a message about official channels. Gemini Spark connects to services through authorized integrations and MCP, not through workarounds. It's the same principle I defend when comparing WhatsApp Business App with the Official API: serious automation runs on sanctioned channels, not on loopholes that can close at any moment.

My recommendation for those managing technology in a company is simple: don't chase Gemini Spark now, but start designing your processes as if an agent will execute them in a few months. Those who document their own tasks well today will delegate with a click tomorrow.

Risks and limits before delegating tasks

Enthusiasm is easy; caution is what separates mature adoption from headaches. Gemini Spark raises legitimate questions worth facing before trusting it.

The first is permission and access. An agent that reads your Gmail, your credit card statement, and your documents concentrates a sensitive volume of data. Google states that Gemini Spark asks for explicit authorization before high-risk actions — spending money or sending emails on your behalf. It's a necessary control, but it only works if you take it seriously and don't click "allow" on autopilot.

The second is agent error. AI models still make mistakes, and an agent that acts alone can err without anyone watching. A wrong email summary is annoying; a wrong payment is expensive. That's why the prudent recommendation is to start with reversible tasks — summaries, drafts, organization — and only later release those that involve money.

The third is platform dependency. The more your routine depends on Gemini Spark, the more you depend on Google: its prices, its terms, its availability in your country. For a company, this is a strategic decision, not just a technical one. It's worth adopting — as long as you have a clear plan B.

Conclusion: the beginning of the agent era

Gemini Spark is not a new feature of Gemini; it's the bet that the next chapter of AI will not be conversing with models, but delegating tasks to them. Announced at Google I/O 2026, still restricted to the US and paid plans, it's more a signal of direction than a product ready for the Brazilian market.

The smart move now is not to wait for Gemini Spark to arrive. It's to understand the agentic logic and prepare your processes for it — because Google, and the competition, are all heading in the same direction. If you want to discuss how to apply automation and AI responsibly in your operation, talk to the Agathas Web team: we help separate hype from what delivers results.