Gemini Spark: How Google's 24/7 Cloud Agent Actually Works

Most coverage of Gemini Spark describes it as an AI assistant that lives on your phone. It doesn't. Spark runs on dedicated Google Cloud virtual machines that keep working when your laptop is closed and your phone is in your pocket, and that single architectural choice explains almost every interesting thing about how the product behaves, what it costs, and why its privacy story is so contested.

Google announced the agent at I/O 2026 on May 19, alongside a refreshed Antigravity builder app and a new Managed Agents API. The pitch is straightforward. The execution, less so.

The Misconception That Trips Up Almost Every Early Review

Spark is not a chatbot with extra tools bolted on. According to Google Cloud's official I/O 2026 recap, Spark is positioned as a "24/7 personal AI agent that helps you work more efficiently by autonomously taking action on your behalf, under your direction," aimed at Gemini Enterprise and Workspace customers. The phrase that matters there is autonomously taking action. Spark is allowed to do things while you are not watching.

That is the meaningful break from ChatGPT, Claude, and every previous Gemini surface. You open a chatbot, ask a question, and close the tab. You hand Spark a goal, walk away, and it keeps running on Google's servers. As BuildFastWithAI's technical breakdown lays out, Spark executes on dedicated Google Cloud VMs rather than on the user's device, which is what makes the 24/7 framing literal rather than marketing.

Think of it less like Siri and more like a contractor with a key to your house. They can keep working when you leave. That is the appeal. That is also the problem.

What Sits Under the Hood

The model behind the agent is Gemini 3.5 Flash. Google Cloud's announcement describes it as delivering "exceptional computational efficiency" for agent orchestration, which is corporate phrasing for cheap enough to run continuously without bankrupting the unit economics. DataCamp's analysis cites Google's own claim that 3.5 Flash "outperforms Gemini 3.1 Pro on agentic and coding benchmarks while running four times faster than competitors," though Google has not published the underlying test set.

Spark's reach inside Google's own products is where it has the clearest structural advantage. Because it connects to Gmail, Docs, and Calendar through native APIs rather than scraping the screen, DataCamp notes the integration is more reliable but also more tightly bounded to Google's own services. Outside that walled garden, the agent reaches into the open web through Model Context Protocol connectors. At launch, Spark supports Canva, OpenTable, and Instacart as third-party MCP integrations, with more partners being added.

Thirty-plus connectors total, by BuildFastWithAI's count. Claude Cowork ships with access to more than 2,300 MCP servers. That is not a small gap.

For builders, Google also opened a Managed Agents API on its Agent Platform, which the official Google Cloud post says lets developers spin up custom agents inside "secure, Google-hosted remote environments using a single API call." Spark itself is built on the same Antigravity agent infrastructure that powers the new Antigravity 2.0 desktop builder.

The Spec That Nobody Is Talking About: Where Your Data Sleeps

Here is the detail buried in the marketing. Because Spark runs on Google's servers rather than locally, every email it drafts, every calendar invite it reads, and every Doc it touches passes through, and is processed by, Google-hosted infrastructure. BuildFastWithAI frames this bluntly: Claude Cowork's local-first architecture gives it a stronger privacy posture than Spark's cloud-VM model, and the gap is not closed by any setting Google currently exposes.

This matters because of timing. The Thele v. Google LLC proposed class action, filed in San Jose federal court in November 2025, alleges that Google quietly enabled Gemini across Gmail, Chat, and Meet accounts in October 2025 without user consent. That suit is the live legal context every Spark privacy claim now sits inside. The EU AI Act's consumer-facing AI agent obligations also take effect on August 2, 2026, and as of the keynote, Google had not published a Spark-specific privacy policy.

The most candid privacy disclosure is from Google itself. Memeburn's reporting on the onboarding flow notes Google's own warning that Spark "may do things like share your info or make purchases without asking" in some situations, and the company advises active user supervision. Google explicitly states Spark is not suitable for medical, legal, or financial decisions. Read that sentence twice. The vendor is telling you, in onboarding, that the autonomous agent may purchase things on your behalf without confirming first.

That is the Gemini Spark privacy story in one line. The product is shipped with a disclaimer most reviewers have not quoted in full.

Gemini Spark vs ChatGPT Agent: Where Each One Wins

The head-to-head with OpenAI's offering is less close than the marketing implies, and the strengths are clearly partitioned. According to BuildFastWithAI's competitive analysis, ChatGPT Agent leads on computer-use and desktop automation, measured by OSWorld benchmark scores, while Spark leads on zero-friction access to Google Workspace.

The practical translation:

• If your workflow lives in Gmail, Docs, Sheets, and Calendar, Spark's API-level integration is faster, more reliable, and less prone to the screen-reading errors that plague computer-use agents.
• If your workflow involves driving arbitrary desktop apps, filling forms in legacy web tools, or operating software that does not expose a clean API, ChatGPT Agent's OSWorld lead is the more honest benchmark.
• If you care most about a large connector library, Claude Cowork's 2,300+ MCP servers dwarf both.

Pricing is where the asymmetry is harshest. BuildFastWithAI pegs Spark's entry point at $100 per month through Google AI Ultra, while Claude Cowork is included in Claude Pro at $20 per month at launch. Sources disagree on the upper tier: Memeburn and other outlets cite a $250 per month figure, suggesting Google operates multiple sub-tiers under the AI Ultra umbrella rather than a single price. For context on the entry plan, see our prior analysis of Google's $100 AI Ultra plan.

A five-times-pricier product can absolutely be the right answer for an enterprise Workspace customer whose employees already operate inside Google's stack all day. For an individual, the math is harder to justify, and Google has not announced any free-tier timeline.

How Google Always-On AI Agent Behavior Actually Plays Out

The "24/7" claim is not metaphor. BuildFastWithAI confirms Spark persists even when the user's phone is locked or laptop is closed, because the agent state lives on a remote VM rather than on a device. That is the difference between a cron job you have to babysit and an assistant that keeps watching your inbox while you sleep.

The control layer Google built around this is one explicit pillar of the announcement. Google Cloud's I/O post lists "Maintain complete control" as a design pillar, requiring high-risk actions such as sending emails to receive explicit user approval before execution. Spark works in the background across Workspace, custom connectors, and the open web, but the gating is supposed to catch the irreversible steps.

The friction is exactly where you would expect it. An agent that confirms every email send is one that cannot really run unattended. An agent that does not confirm is one whose own vendor warns it might make purchases without asking. Google has decided the line sits somewhere between those two, and the location of that line will probably move several times during the beta.

The Google AI Ultra agent is currently a US-only beta limited to AI Ultra subscribers, per DataCamp, with no announced international rollout. Memeburn reports that trusted-tester access began the week of May 19, 2026, with broader U.S. AI Ultra subscriber access following the subsequent week. A standalone Mac app with local file access is planned for summer 2026, though Google has not confirmed a release date.

The 19-Billion-Tokens-Per-Minute Argument

Google's strongest non-feature pitch is scale. BuildFastWithAI's writeup cites a Google figure of 19 billion AI tokens per minute processed across its products, presented as the infrastructure advantage that makes a 24/7 background agent economically viable in the first place.

The argument is not that 3.5 Flash is the smartest model. The argument is that Google can afford to keep it running for you indefinitely, on a Google-hosted VM, in a way smaller competitors cannot match without exploding their unit economics. That is a defensible claim, and it is also why Spark's design choices look the way they do. Server-side persistence is cheap for Google in a way it is not cheap for Anthropic or OpenAI yet.

The trade is real. You get an agent that never sleeps. Google gets a much richer telemetry stream about how knowledge work actually happens, on top of a Workspace dataset it already owned. That second part is the half of the deal most marketing copy leaves out, and it is the half a future regulator is most likely to probe.

What This Means Before You Sign Up

If you are deciding whether to pay for Spark right now, the calculus is narrower than the I/O keynote made it sound. Three concrete checks are worth running before you commit:

1. Does your work actually live in Workspace? If most of your day is Gmail, Docs, Sheets, and Calendar, the API-level integration is where Spark earns its price premium over ChatGPT Agent or Claude Cowork. If not, the value proposition collapses fast.
2. Are you comfortable with the onboarding warning? Google's disclosure that the agent may share information or make purchases without confirmation is the single most important sentence in the entire product, and it is not negotiable through settings as of launch.
3. Can you wait six months? A Mac app with local file access is on the roadmap for summer 2026 per Memeburn, the EU AI Act's agent rules land August 2, and the Thele class-action discovery process will surface concrete information about how Google handles Gemini data. Each of those will change the product or its risk profile.

The interesting question is not whether Spark works. It does. The interesting question is whether "a contractor with a key to your house, employed by the company that owns the house" is the right shape for an autonomous agent to take, or whether the local-first architecture Anthropic is betting on will turn out to be the more durable model. That answer arrives over the next eighteen months, not at I/O. For more on how the rest of the I/O announcements fit together, see our Google I/O 2026 preview coverage and the broader AI Industry articles archive.