Gemini Spark
Last reviewed
Jun 3, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,485 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,485 words
Add missing citations, update stale details, or suggest a clearer explanation.
Gemini Spark is a personal AI agent from Google that runs continuously in the background and takes action on a user's behalf across their digital life. Google announced it on May 19, 2026 during the opening keynote of Google I/O 2026, describing it as a "24/7 personal AI agent" that lives inside the Gemini app.[1][2] Unlike a chatbot that waits for prompts and answers in a single session, Spark is meant to keep working over hours or days, monitoring information, organizing data, and carrying out multi-step tasks while a person's phone or laptop is closed or even turned off. It is powered by Gemini 3.5 Flash and Google's Antigravity agent platform, and it launched in beta for subscribers on the Google AI Ultra plan in the United States.[3][4]
Spark sits in the broader category of AI agents: systems that do not just answer questions but plan and execute work on their own. Google frames it as the consumer face of the agentic research it showed across the rest of I/O 2026. Sundar Pichai, in his keynote, called it "your personal AI agent in Gemini app that helps you navigate your digital life, taking action on your behalf and under your direction."[2] The "under your direction" part is doing a lot of work in that sentence, and Google repeats it often. Spark is opt-in, it can be pointed at specific apps and data, and Google says it asks for permission before anything consequential.
The thing that separates Spark from the regular Gemini assistant is persistence. A normal assistant session ends when you close the app. Spark keeps running on dedicated virtual machines in Google Cloud, so it can handle long-horizon jobs in the background without tying up your device or draining its battery.[1][4] That design choice is what lets Google market it as a "24/7" agent rather than a faster autocomplete. Google's official product page puts it plainly: Spark "works autonomously in the background 24/7, even if your phone and laptop are turned off."[4]
Spark is not a separate model. It is an agentic product built on top of Gemini, specifically the Gemini 3.5 Flash model that Google unveiled at the same event, paired with the orchestration layer from Antigravity, Google's agent development platform.[3][4] Flash is the faster, lower-cost tier of the Gemini 3.5 family, which matters for an agent expected to run continuously and cheaply rather than burn premium compute on every step. Antigravity supplies the harness that lets the model take actions, hold state across a long task, and coordinate tool use.
Inside the Gemini app, Spark appears as a distinct mode you can switch into rather than a wholesale replacement for the assistant. You still chat with it in plain language. The difference is intent: a question gets answered, but a task gets handed off to Spark to work on over time. That split, between "answer this" and "go do this," is exactly the design decision that drew the most criticism after launch (more on that below).
Google organizes Spark's behavior around three ideas it calls Tasks, Schedules, and Skills.[4]
| Feature | What it does |
|---|---|
| Tasks | One-off jobs Spark researches and completes, such as finding products on sale or compiling a list of weekend events |
| Schedules | Recurring or condition-triggered routines, for example checking a credit card statement for hidden fees every month |
| Skills | Custom automations Spark learns from how an individual works, becoming more tailored over time |
In practice, the early examples Google and reviewers describe include pulling deadlines out of a Gmail inbox and sending them back as a summary, condensing long email threads, drafting reports from meeting notes, watching for price drops on an expensive item, and surfacing coupons that match a shopping list.[1][5] Spark can read from a user's Gmail, Docs, Sheets, and Slides to draft communications and assemble outputs, and Google says it can browse the open web directly through Chrome.[1]
The integration story extends past Google's own apps. At launch Spark connected to Google Workspace (Gmail, Calendar, Drive, Docs, Sheets, Slides, Maps, and YouTube) plus a handful of outside services including Canva, OpenTable, and Instacart, with more partners promised.[4][5] For everything else, Google leans on the Model Context Protocol, the open standard for wiring agents to third-party tools, which it said would expand Spark's reach in the weeks after launch.[2][5]
A few interaction details set it apart. Spark gets its own dedicated Gmail address, so you can email the agent a task the way you would forward something to an assistant.[1] On Android, a feature Google calls Halo gives you a live view of what the agent is doing and the updates it sends back. And later in the summer of 2026, Google planned to let Spark operate as an agentic browsing companion inside Chrome and to ship a desktop app.[2][5]
On the safety side, Spark is built to check in before it does anything with real stakes. Google says it "will ask you first before performing high-stakes actions like spending money or sending emails."[5] The agent is fully opt-in, you decide which apps it can touch, and Google stresses it "does not read your emails indiscriminately."[4] A future roadmap item is letting users authorize payments with budget and merchant limits, so Spark could buy things within guardrails the user sets in advance.[1]
Spark launched as an early beta, and Google was explicit that it is "very early in its product journey."[1] Access rolled out in stages: trusted testers got it during the week of the May 19 keynote, and U.S. subscribers on the Google AI Ultra plan received beta access the following week.[1][2] Google's product page lists eligibility as Google AI Ultra subscribers over 18 in the United States, along with select business users.[4] Google AI Ultra is the company's top consumer subscription, priced at $250 per month, which makes Spark a premium feature at launch rather than something available to the free Gemini tier or the cheaper AI Pro plan.[3] Whether it reaches lower tiers later was left unconfirmed.
Coverage of Spark split along a familiar line for agentic products: the capability is impressive, the readiness and the packaging are the open questions.
TechCrunch's Sarah Perez tested the beta across six everyday jobs, from finding drugstore deals to planning a weekend, and came away genuinely positive on the research and curation side. Spark "told me exactly what products were on sale that matched my needs, and suggested coupons to clip," and its packing-list suggestions were "spot-on."[5] Her verdict, captured in the headline, was that the assistant is "actually pretty useful."
The criticism was just as pointed. The same review flagged real gaps: Spark could not use Google Keep, an odd omission for a productivity tool, and it could not yet handle common third-party jobs like booking a regular dinner through Resy or hunting flight deals on a preferred site.[5] Outputs were sometimes incomplete, returning four newsletter summaries when five were asked for, or skipping the cost and date details that would make a list of activities actually actionable.
Perez's sharpest objection was about branding. Her argument was that there is "no need for this to be a stand-alone product with a different branding," and that asking users to "determine whether something is a question or a task" before they ask just adds confusion to what should be one seamless assistant.[5] In her framing Spark "doesn't deserve to have its own brand" even though it works well.
Beyond usability, an always-on agent that reads your mail and acts on your accounts raises obvious privacy and oversight questions, which several outlets noted alongside the unconfirmed long-term pricing.[5] The agent does pause for approval before major actions, but as critics pointed out, the line between "major" and routine is fuzzy: a purchase clearly needs a check, but does every email? Where that boundary sits, and how it shifts as Google grows confident in the system, is likely to be one of the more closely watched aspects of the product as it expands.