Gemini Spark: A beginner’s guide to Google’s AI agent

Miss the trigger, and you get a one-off instead of an automation. Miss the destination and the work vanishes into a chat reply you will never find again. The condition is optional, but it is what stops Spark from burying you in noise.

You need less context than you think. In the APAC briefing, Coimbra asked Spark to project-manage a studio renovation and deliberately gave it almost nothing. No link to the file. He even used the wrong word.

“In fact, I called it a remodel. The actual PDF says renovation. It doesn’t actually say remodel anywhere,” he said. “So my prompt wasn’t even a good prompt, but it immediately found the file.”

Spark searched the Drive, found the renovation plan, and rebuilt the entire schedule a month forward, creating a Google Task for each step with details and times, plus calendar events for the major milestones.

The lesson is to describe the outcome you want rather than laboriously assembling the inputs. Spark will go looking. As Coimbra put it, it “will automatically understand that you’re a Google user and it’s gonna go find what context it can.”

Level one: your first task

Start with something that has an obvious destination and an obvious trigger. A tracker is ideal.

The demo Google used was an RSVP tracker. The instruction was, in effect, create a spreadsheet to track workshop RSVPs, and log each one as the emails arrive.

Spark immediately began searching Gmail and Drive for context about the workshop without being told to. It then built the spreadsheet, found previous RSVPs already sitting in the inbox and preloaded them, and set up a scheduled action to fire whenever a new RSVP email arrives, extracting the name, email address, date and status, then appending a row.

When a test email arrived, Spark woke up on its own, read it, and updated the sheet.

It also mangled the formatting, which Coimbra logged as a bug on the spot. Worth knowing before you rely on the output looking presentable.

Level two: make it proactive

Most AI tools can already run something daily or weekly. “Give me a news update,” or “Summarise the emails I got about this project.” Spark goes further, because you can point it at a subject rather than a time.

Tell Spark to keep track of a project and it works out that it should wake up hourly, or the moment an email on that topic arrives, and then go and see what has changed. If you use Google Chat, it can read what people are saying across different chat rooms, pull together an update, and maintain a tracker for you in a Google Sheet.

This is the difference between a reminder and an agent. A reminder tells you to check. An agent checks.

Level three: teach it a skill

The easiest skill to start with is one Google has already built for you. Open the skills section and look under recommended for Match My Writing Style.

The first time you use it, it reads through your emails and documents and constructs a custom skill around how you actually write. From then on, anything Spark drafts for you sounds more like you.

You can also build one from scratch by describing it, or by pointing Spark at examples. Google’s own suggested prompt for this is worth copying directly:

“Read through the last 50 emails that I wrote and turn it into a style guide for how I write emails. Turn that into a skill that gets used every time I ask you to draft emails for me. Call that skill ‘ghostwriter’.”

Note the ‘ghostwriter’ instruction. Naming a skill means you can invoke it deliberately later.

Level four: chain a skill to a schedule

This is where Spark stops being a novelty, and the best example came from Coimbra describing his own setup rather than from any slide.

He built a skill that reads through all his Google Chats and emails from the past couple of days and produces a list of everything he has promised to do, plus everything other people have promised him. Then he built a schedule that runs that skill automatically and emails him the result.

“I’m constantly saying yes to things, and yes, I’ll do that, and yes, that sounds good,” he said, “and then I lose track of them.”

The email he received on the morning of the briefing ran to five pages.

The pattern generalises. A skill defines how to do something well. A schedule defines when. Build the skill once, attach a schedule, and the job runs without you ever opening Spark again.

Letting Spark use your browser

Spark can drive a browser to complete tasks on websites that have no Google integration at all. In the demo, it was asked to go to Class Bento, a platform for listing local courses, and post a workshop listing for around ten people per session on Wednesday nights.

Two things are worth knowing before you try this.

• It will ask permission first. Spark researched the task by digging through Google Keep and Drive, then stopped and asked before opening a browser at all.
• You get to choose whose browser. You can let Spark drive Chrome on your own machine, where you will watch the cursor move by itself, or you can hand the job to its own browser running on a virtual machine. Pick the virtual machine if you want to close your laptop and walk away. Pick your own Chrome if the site needs you to be logged in already.

In the demo, Spark got most of the way through the listing, realised it did not have pricing information, and came back to ask rather than inventing an answer. That is the behaviour you want, but it does mean browser tasks are not always set-and-forget.

Ready-made prompts to copy

Google has published a set of starter prompts. These are the ones worth stealing first, and each is a clean illustration of the four-part pattern above:

What Spark cannot do yet

Four limitations to know before you build anything important on it:

Staying in control

Spark only does what you have told it to do, and it pauses for approval before anything sensitive. High-stakes actions such as sending emails, moving files or making payments always stop for your explicit sign-off.

Beyond that first line of defence, Google runs systems that monitor Spark’s activity to check it is staying aligned with what you originally asked for, and the underlying models are trained for safety and continuously evaluated.

Google is also deliberately stopping short of letting Spark spend your money. Coimbra used the Instacart integration as the illustration: Spark can fill your basket, but “it will only add things to your cart, and then you will have to” complete the purchase yourself.

This matters because agents have gone wrong elsewhere. In January 2026, shortly after Anthropic launched Claude Cowork, venture capital founder Nick Davidov asked it to tidy his wife’s desktop and granted permission only for temporary Office files. It deleted a folder holding fifteen years of family photos, somewhere between 15,000 and 27,000 files, bypassing the Trash entirely. OWASP’s Top 10 for Agentic Applications, published in December 2025, ranks agent goal hijacking as the number one risk in the category.

“I’m not claiming that this is perfect. It’s still a beta product, and so you should keep it under supervision if you’re asking it to do anything that’s in any way sensitive to you,” Coimbra advised.

One refinement worth waiting for, as right now, what counts as sensitive is a general standard, with certain actions such as sending an email always requiring confirmation. Coimbra wants to loosen it so you can eventually tell Spark it no longer needs to ask permission for particular types of action.

Spark versus the alternatives

When we asked Google to spell out the differences between ChatGPT Work and Gemini Spark, all they would offer was a restatement of Spark’s five selling points.

From what we have seen, the practical distinction lies in Spark being always on and deeply integrated with the rest of Google’s apps. If you already live in Gmail, Drive and Calendar, that integration is the whole argument, and Spark is the agentic AI of choice.

What is coming next

Spark has moved quickly since I/O 2026. Google has shipped custom MCP integrations, new connected apps, desktop availability on Mac, and the ability to track topics in real time on the web. Chrome Auto Browse, the browser automation described above, went from a preview in the briefing to a public release within days.

Two things worth watching. Spark is coming to your texting app of choice and to email, so you will be able to forward it a message full of to-dos and have it start working through them. And Google is rolling out sub-agents, or as Coimbra kept calling them, baby sparks. Spark already spins these up automatically for web research, and the longer-term version involves building out an agent and letting it interact with you and with other people.

Note that Spark is not yet in final release and changing quickly, so check the limitations section for what has since been fixed