Google has begun rolling out Gemini Spark, an always-on AI agent that draws on a user's Gmail, Google Docs, Calendar, and Photos to take action across apps, with early access available on the $99/month AI Ultra plan. The product is Google's most aggressive answer yet to the agent race, pitched as a single interface that can operate external apps and, over time, the user's computer. Editor-at-Large David Pierce got early access and described Spark in The Verge as both the most impressive and the most invasive AI experience he has had.
Spark sits on top of what Google calls Personal Intelligence, the layer that lets Gemini reach across a user's entire Google footprint. In Pierce's testing, simple agentic jobs worked: Spark combed his Gmail inbox and produced an organized document of subscriptions to cancel, complete with unsubscribe links, and pulled stale tasks out of his Google Docs. Those are the kinds of tasks OpenAI's and Anthropic's agents have been pitching for the better part of a year.
The harder test was a trip-planning prompt: a weekend in Hershey, PA with a wife, two kids, and a dog, for the weekend of July 18. Spark returned a couple thousand words of itinerary covering Friday, July 17 through Sunday, July 19, 2026, organized into a Google Doc and shared back to the user. The level of personalization went well beyond what was in the prompt.
Key facts
- 01Gemini Spark is rolling out to Google's $99/month AI Ultra plan as an always-on agent designed to operate apps and, eventually, the user's computer.
- 02In a trip-planning test, Spark produced a roughly two-thousand-word itinerary for a weekend in Hershey, PA from July 17–19, 2026.
- 03Spark pulled personal details Google was never explicitly given for the prompt, including the user's home address, dog's name, spouse's food preferences, and kids' ages.
- 04Spark successfully drafted a Google Doc and emailed it to the user's wife, but was blocked by Airbnb when it tried to complete a booking.
- 05Google's data advantage — Gmail, Docs, Photos, Calendar, Search — feeds Spark via the Google Personal Intelligence layer.
Spark inserted driving directions from the user's home, an address it pulled from his Google account rather than the prompt. It listed hotel options with pet fees, named the family dog Frida — a detail Pierce believes Spark surfaced from vet emails — and noted that one-year-old Lewis would enter Hershey Park free while three-year-old Arthur would need a ticket. Spark also scheduled nap time for 1:30 PM and worked around the wife's dislike of onions and scallions.
“I have created a comprehensive, family-friendly, and dog-friendly weekend itinerary for your trip to Hershey, PA, from Friday, July 17 to Sunday, July 19, 2026.”— Gemini Spark, Google AI agent (in-product output)
It identified a Thomas Rhett and Niall Horan concert on Saturday night, almost certainly from a Ticketmaster confirmation in Gmail, and noted that parking was already covered by the tickets. When Pierce added mid-conversation that his parents were coming to babysit, Spark addressed them by name and switched the lodging recommendation from a hotel to an Airbnb. Asked to share the plan with his wife Anna, Spark found her email, attached the document, drafted a note, and sent it.
Spark hit a wall at the booking stage. After being granted permission to interact with websites, it navigated to Airbnb and was blocked: "Due to security and authentication policies on Airbnb, I am unable to log in, handle payment, or complete bookings directly on your behalf." Instead, it surfaced available listings for the relevant dates and listed the information Pierce would need to complete the booking himself. The block appears to be Airbnb's, not Google's — a preview of the friction agents will encounter as they try to act inside other companies' authenticated experiences.
Spark's broader ambition is to be the interface for everything else on a user's device. One way to describe it, Pierce writes, is "OpenClaw with better internet access" — a reference to Anthropic's computer-use agent, but with the entirety of a user's Google graph already loaded behind it. That data advantage is the whole game: OpenAI and Anthropic are still working to accumulate the personal context that Google has been collecting for over a decade.
“What Spark did feels sort of magical, and very invasive.”— David Pierce, Editor-at-Large, The Verge
The trade-off is the obvious one. Spark is more useful precisely because it knows more, and it knows more because Google already has the emails, the photos, the calendar entries, and the search history. Pierce's discomfort wasn't that Google had the data — most users intellectually accept that — but that Spark treated the data as raw material to be mined and surfaced rather than something held quietly in the background. A vet email becomes a dog's name in a hotel recommendation; a Ticketmaster receipt becomes a Saturday night plan.
Spark's failure modes will get scrutiny as the rollout widens. Airbnb-style authentication blocks are one limit; the bigger questions are whether Spark will correctly handle ambiguous instructions, what happens when it acts on stale or wrong data pulled from years-old emails, and how it behaves when a user wants some personal context excluded from a given task. Google has not publicly detailed Spark's controls for scoping which data sources it can read on a per-task basis.
Spark is the clearest demonstration so far that the agent era will be won on data, not on model quality alone. A frontier model with no context produces generic itineraries; a competent model with full read access to a decade of Gmail produces something that feels like a human assistant. That asymmetry is why Google can charge $99/month for AI Ultra and why every competitor without a comparable consumer graph — OpenAI, Anthropic, Perplexity — is racing to build memory layers, browser integrations, and connectors. Spark is the product Google's competitors have been quietly afraid of, and the one that will force the rest of the industry to decide how much of a user's life it wants to ingest to keep up.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
