Google's Spark Wants to Be Your 24/7 Coworker

A Sunday evening problem

Picture a Sunday evening. You remember three loose ends you meant to handle before Monday. Book a table for a client dinner. Dig through your inbox for the vendor quote you swore you saved somewhere. Start a spreadsheet comparing two software options your team needs to pick between. None of them are hard. Together, they're enough to ruin your night.

At Google I/O this past Tuesday, Google introduced something built for that exact moment. It's called Gemini Spark. ¹ The pitch is simple. You hand off tasks like those in plain language, and Spark works on them in the background while you go finish dinner.

That's the kind of promise we've heard a hundred times. The reason this one matters is the plumbing underneath it.

What Google actually shipped

Gemini Spark is Google's new personal AI agent, baked into the Gemini app and rolling out to "trusted testers" this week. ² Beta access opens for US Google AI Ultra subscribers next week. The Ultra plan dropped to $99.99 a month at I/O, down from $249.99, almost certainly so Spark has a real shot at landing with paying customers rather than just developers. ³

Three things make Spark different from talking to a chatbot.

It runs on its own virtual machine in Google Cloud, so it keeps working when you close your laptop. ¹ That sounds like a small detail, but it's the whole game. Previous AI assistants were essentially conversations. Spark is a worker that doesn't need you sitting there to function.

It's powered by Gemini 3.5 paired with something Google calls the Antigravity harness, which is basically the operating system for agents. ¹ Antigravity handles the boring but critical stuff: scheduling, tool calls, sandboxing, credential management, and the kind of background tasks that used to fall apart halfway through.

And it connects to your other apps through MCP, the open standard Anthropic introduced last year. ⁴ Spark can already talk to Workspace, and Google announced new MCP connections to Canva, OpenTable, and Instacart at the keynote, with more coming soon.

Why MCP is the quiet story here

If you've followed this blog, you've heard MCP mentioned a few times. It stands for Model Context Protocol, and it's basically a universal adapter that lets an AI agent talk to any tool that speaks the protocol. ⁴

What's interesting about Spark is that Google, after years of running its own custom integration plumbing, just adopted MCP wholesale. That's a big deal. It means a small business owner could, in theory, build one Canva integration and have it work for Claude, ChatGPT, and now Spark. The agent ecosystem is settling on a common language, which makes it dramatically easier to build for and dramatically less risky to invest time learning.

For IT folks, this is the moment where MCP stops being a curiosity and starts being something you'll see in vendor pitches and security reviews. If you haven't poked at the protocol yet, this is a reasonable week to start.

What this means if you work in AI or IT

Three honest takeaways for the people reading this blog.

If your role involves automating workflows or managing internal tools, the bar for what "an assistant" means just moved. A request like "watch my inbox for invoices from these three vendors and drop them in the right Drive folder" used to require Zapier, a custom webhook, and a quiet Sunday. With Spark and similar agents from Anthropic and OpenAI, that's increasingly something an employee without an engineering background can set up themselves in plain English. The question shifts from "can we automate this?" to "who gets to authorize the agent, and what guardrails do we put around it?"

If you handle security or compliance, this is the moment to get ahead of the conversation. Agents that live in the cloud, act on behalf of users, and have access to email and Workspace and outside apps will absolutely show up in your environment whether you're ready or not. The questions to start asking now: what scopes can the agent request, how do we audit what it did, and what happens when an employee leaves with a Spark instance still running in their account.

If you're a working professional worried about AI taking your job, the honest read is that Spark, like every other current agent, still needs someone steering it. The 24/7 nature is the headline, but real productivity comes from people who can frame the right task, judge the output, and step in when the agent gets confused. Those skills are not going anywhere. They might actually become the most valuable ones you have.

The catch, because there's always one

Spark is impressive and limited at the same time. It's available in the US only at launch, restricted to Ultra subscribers, and Google is being unusually careful about the rollout, starting with trusted testers and slowly expanding. ⁵ The Antigravity harness is brand new. Complex agent tasks still fail in surprising ways across the industry, and Google's own demo videos at I/O were carefully curated.

That doesn't mean sitting on the sidelines is the right move. It means treating Spark, and Claude's agent features, and ChatGPT's operator as previews of the workflow we're all going to be using in a year or two. The agents that ship today are weaker than the ones that will ship six months from now, and the people who have already wrapped their head around how to work with them will have a real edge when the dust settles.

Where this leaves you

The agent era stopped being a concept this week and started being a product you can pay $99 a month for. ³ That's not a reason to panic, and it's not a reason to ignore the news cycle either. The most useful posture right now is curious and unhurried. Try Spark or one of its competitors when you get access. Notice what makes it useful, and notice where it falls down. Talk to your team about what an agent should and shouldn't do on your behalf.

The professionals who are going to thrive in the next phase of this aren't the ones who memorize every model name. They're the ones who understand how to hand off work, how to verify it, and how to keep humans in the driver's seat. That's a skill set you already have. Spark just gave you a new place to use it.