Gemini Spark: Google's 24/7 AI Agent That Works While You Sleep

Google announced a lot at I/O 2026. Most of it was incremental. Gemini Spark is not.

Spark is a persistent AI agent that runs on Google's servers around the clock. When your laptop is closed, your phone is in a drawer, and you're nowhere near a keyboard, Spark can be drafting responses to your email, monitoring your inbox, updating documents, or completing the calendar tasks you assigned it. The model doesn't need you present to operate.

This is a structural departure from how current chatbots work — and it's worth understanding exactly what changed.

How Spark Works

Gemini Spark runs on Gemini 3.5 Flash, Google's fastest frontier model, on dedicated virtual machines in Google Cloud. That server-side execution is the key architectural fact: the agent isn't tied to a browser tab or a local process that closes when you leave. It runs independently of your device entirely.

The integration layer uses Model Context Protocol (MCP), the same open standard Anthropic and other labs have adopted. This lets Spark connect to Google's own products — Gmail, Docs, Calendar, Sheets — and to any third-party service with an MCP server. The MCP ecosystem is expanding quickly across enterprise software.

On Android, Google introduced a visible notification band called "Halo" showing which background tasks Spark is running at any given moment. The stated purpose is transparency: you should know what the agent is doing on your behalf, even when you didn't prompt it to act.

What Spark Can Do at Launch

The four core capabilities as of release week:

Draft email responses by pulling from existing messages, documents, and spreadsheets
Monitor inboxes for specific message types or contacts
Browse the web via Chrome for research and task-dependent lookups
Execute multi-step workflows across Gmail, Docs, Calendar, and Sheets simultaneously

The multi-step execution is what separates Spark from a smarter autocomplete. Josh Woodward, VP of Google Labs, described a representative example: ask Spark to draft a client proposal, and it pulls contract history from Gmail, pricing from a Sheets file, and generates a Docs draft without requiring you to direct each step individually. That's the workflow the architecture is designed for.

Whether it delivers that reliably in production is a different question. Architecturally, the system supports it.

Pricing and How to Access It

Spark is rolling out to Google AI Ultra subscribers in the US. Ultra is currently $100/month, down from $250 at I/O 2026. That positions it in direct competition with OpenAI's Pro plan ($200/month) and Anthropic's Max plan ($100/month). Initial availability is limited to trusted testers; broader Beta access to Ultra subscribers was promised for the week following the announcement, which puts it in users' hands by late May 2026.

What Sets It Apart from ChatGPT and Claude

ChatGPT, Claude, and standard Gemini all operate reactively: you open a session, type a message, get a response. The model is idle when you're not using it.

Spark's persistent background execution puts it in a different category. The closest analogs are OpenAI's Operator, which completes web tasks autonomously, and Anthropic's "dreaming" capability (also announced in late May 2026), which lets Claude agents review their own behavior between sessions. Neither is a direct match — Spark is designed for continuous background operation, not on-demand task bursts.

The real competitive differentiator is Google's existing data access. Spark connects to the Gmail and Workspace data Google already holds. For anyone running their work life through Google's suite, that's structural depth no other AI company can replicate without first convincing users to migrate years of email, documents, and calendar history. An earlier piece here covered Google's strategic pivot toward agents over traditional chat — Spark is the first concrete product from that shift.

What You're Agreeing To

Enabling Spark means granting a persistent AI system autonomous access to your email, documents, and calendar. Google's infrastructure already processes this data for its own products. Spark adds an agent layer that can act on that data without requiring you to initiate each action explicitly.

As of launch week, Google hasn't published a separate Spark privacy policy. The existing Google Workspace terms predate autonomous agent execution as a product reality. For a full look at how the major AI providers handle account data, this comparison of AI assistant privacy models covers the relevant policies in detail.

The accountability question for background agents remains open at launch: when Spark takes an action with downstream consequences, the Halo notification system shows what's running but doesn't confirm each step before execution. That's a deliberate design tradeoff — one worth evaluating before enabling the agent.

Why This Time Might Be Different

Google isn't the first to build an AI agent with access to your calendar and email. Cortana had it. Google Assistant had it. Siri has had versions of it since 2011. None of them worked well enough for people to rely on them.

What's different in 2026 is model capability. Frontier models can now follow multi-step instructions, maintain goals across long workflows, and produce outputs that don't need to be rewritten before use. The technology finally matches the ambition that motivated all those earlier attempts.

Google's specific advantage is data access; its specific risk is the standard agent deployment problem — reliably doing the right thing when no one is watching. The slow rollout to trusted testers is the right first step. Whether Spark works as advertised across the full range of real workflows is what the late May beta will start to answer.

The About.chat Weekly newsletter covers AI assistants and what the major labs are actually shipping — without the hype or press release rewrites. Free, one email a week. Subscribe here.