You close your laptop. Your AI keeps working. That, in a sentence, is the pitch behind Google’s Gemini Spark — a personal AI agent that runs on dedicated cloud virtual machines, not on your phone or laptop, and continues monitoring your inbox, tracking prices, and running multi-step tasks whether your devices are on or off.
Announced at Google I/O on May 19 and now available in beta to Google AI Ultra subscribers in the US, Spark represents Google’s most aggressive consumer AI play to date. It is also the clearest signal yet of where the industry is heading: away from chatbots that wait for your question, toward agents that act on your behalf without you watching.
What Spark actually is. Gemini Spark is not a new AI model. It is an agent runtime — a persistent system built on Gemini 3.5 Flash and powered by Google’s internal Antigravity platform. When you set up a task in Spark — “summarise my newsletters every Friday,” “track price drops on this product,” “find weekend events near me and add them to my calendar” — the instruction is deployed to a Google Cloud VM where it persists indefinitely. You do not need to keep the Gemini app open, your phone unlocked, or your laptop awake. The agent runs in Google’s data centre, checks in on your accounts when scheduled, and surfaces results to you via notifications.
TechCrunch’s Sarah Perez, who tested Spark before its public rollout and [published her hands-on review on May 30](https://techcrunch.com/2026/05/30/i-put-googles-24-7-ai-assistant-gemini-spark-to-work-and-its-actually-pretty-useful/), described the experience as “fairly useful” but noted that the product suffers from branding confusion and integration gaps. Spark lives as a separate toggle within the Gemini app — distinct from the chatbot interface — which means you have to decide in advance whether you are asking a question or assigning a task. “I don’t want to carry the mental load of trying to determine whether something is a question or a task,” Perez wrote.
The cloud-VM bet. The architectural choice to run Spark on Google Cloud rather than on-device is the story’s most important detail. Apple is rumoured to be distilling the multi-trillion-parameter Gemini model to run locally on iPhones for the next-generation Siri. Google, by contrast, is betting that users will accept the latency and privacy trade-offs of cloud execution in exchange for true 24/7 persistence.
The trade-off has real consequences. Spark runs on Gemini 3.5 Flash — a model Google claims delivers 284.2 tokens per second, roughly 4× faster than comparable frontier models — which makes long-horizon agentic tasks responsive enough to be useful. But it also means Spark processes your Gmail, Calendar, and Docs on Google’s servers, not your device. The onboarding screen reportedly includes the disclosure that the agent “may do things without asking,” a phrase that privacy researchers have already flagged as concerning.
What it does well, and where it falls short. Perez’s hands-on test puts real flesh on the architecture. She asked Spark to find weekly deals at a local drugstore, and it successfully identified sale items, clipped coupons via the Walgreens app, and suggested stacking strategies — though one of the promo codes turned out to be invalid. She asked Spark to plan a packing list for a day trip: it checked the weather, gathered event details, and made spot-on suggestions (lawn chairs, sunscreen, an umbrella for possible showers), then offered to put the list in a Google Doc — because Spark cannot use Google Keep, a glaring omission for personal productivity.
Spark also compiled weekend activity suggestions by combining web searches with inbox scans of local newsletters, surfacing events Perez would otherwise have missed (an “Annual Beaver Queen Pageant” raising money for wetland conservation). It tracked price drops on an expensive skincare product, though Perez noted the fortnightly check-in cadence might be too infrequent to catch flash sales. For recurring newsletter digests, Spark returned four articles when asked for five, and the links it generated were Google redirect URLs that required manual navigation.
The integration landscape at launch covers Gmail, Calendar, Docs, Sheets, and Slides natively, with MCP (Model Context Protocol) connections to Canva, OpenTable, and Instacart on day one. Adobe, Samsung, Spotify, and CapCut are expected over the summer.
The agent wars take shape. Spark enters a market that already has Anthropic’s Claude Cowork and OpenAI’s ChatGPT agent. Google’s differentiating advantage is its suite — no other AI company can read your email, check your calendar, edit your documents, and search your photos under one permission model. Its disadvantage is the unbranded nature of the thing: Spark is a separate product name nested inside Gemini, which is itself a brand inside Google’s broader AI portfolio. Perez’s verdict — “there’s no need for this to be a standalone product” — captures a real risk of consumer confusion.
Pricing adds another layer. Spark requires a Google AI Ultra subscription at $99.99 per month (restructured at I/O 2026), which also includes access to Gemini 3.1 Pro, video generation with Veo 3.1, and 20 TB of cloud storage. That positions Spark as a premium add-on in a market where OpenAI’s equivalent agentic features are bundled into ChatGPT Plus at $20 per month.
The deeper question isn’t whether Spark works — it mostly does, with caveats. It’s whether consumers will trust an agent that keeps working while they aren’t watching. Google is betting that convenience wins over caution. The next few months of beta feedback will determine whether that bet pays off.
Sources: [TechCrunch hands-on review](https://techcrunch.com/2026/05/30/i-put-googles-24-7-ai-assistant-gemini-spark-to-work-and-its-actually-pretty-useful/) (Sarah Perez, May 30, 2026); [Build Fast with AI architecture breakdown](https://www.buildfastwithai.com/blogs/gemini-spark-google-ai-agent-how-it-works) (May 20, 2026); [TechCrunch announcement](https://techcrunch.com/2026/05/19/google-introduces-gemini-spark-a-24-7-agentic-assistant-with-gmail-integration/) (May 19, 2026); [AIToolBlaze pricing analysis](https://aitoolblaze.com/blog/gemini-spark-review-worth-100-a-month)