Gemini Spark: Cloud Personal Agent, Google Apps, Ultra Beta

An agent with its own email address

Google announced Gemini Spark at its I/O 2026 developer conference as, in Sundar Pichai's description, a personal AI agent that helps a user navigate their digital life, taking action on their behalf and under their direction. The pitch positions it as a proactive, 24/7 counterpart to the question-and-answer Gemini app. [1]

Spark reaches a user through a dedicated Gmail address rather than a chat window, and Google described the interaction as writing to a colleague. VP Josh Woodward's example was asking Spark for a status update email, which it drafts by pulling facts from the user's mail, docs, sheets and slides. Google also described it monitoring an inbox so a small business does not miss a customer inquiry. [1][2]

A machine that does not need your laptop open

Spark runs on Gemini 3.5 combined with an agent harness from Google Antigravity, executing on dedicated virtual machines inside Google Cloud rather than on the user's own device. Both TechCrunch and The Next Web reported that this is what lets it keep working with a laptop closed or a phone locked. [2][3]

Google's own post describes Spark asking the user first before high-stakes actions such as spending money or sending an email, rather than acting on those without confirmation. Connected services at launch were Gmail, Docs, Slides, Sheets and Chrome browsing, with Model Context Protocol support for reaching outside tools and an Android notification layer, called Halo, for tracking progress on the phone. [1][2][3]

Atlas interpretation: Running the agent on Google's own machine rather than the user's is the part of the design worth noticing, more than the specific list of connected apps. A cloud VM removes the laptop as a dependency and makes multi-hour or multi-day tasks possible, but it also means the account holding Gmail, Docs and Chrome access is one Google controls end to end, with the approval prompt as the main check on what it does unattended. [1]

Google's answer to the coding agents

TechCrunch placed Spark next to Anthropic's Claude Cowork and OpenAI's ChatGPT agent as the same idea aimed at people who do not write code, the way Claude Code and Cursor extended delegated work to programmers first. OpenAI had already tried a version of this for browser tasks with Operator, which drove a browser under human oversight at a fraction of human accuracy on the OSWorld benchmark. [2][3]

Atlas interpretation: The advantage TechCrunch and The Next Web both point to is not a model claim. It is that Google already holds the email, the documents, the browser and the operating system that a personal agent needs permission to touch, so Spark does not have to ask a user to connect those accounts the way a third-party agent would. Whether that access is used carefully is a separate question from whether it exists, and this announcement establishes only the launch, not a track record. [2][3]

Rollout

Google's post describes trusted testers starting the week of the announcement, with a beta opening to U.S. Google AI Ultra subscribers the following week and further app integrations planned over the summer. TechCrunch and The Next Web both confirm the Ultra-only beta timing; The Next Web additionally reported that Google cut the AI Ultra subscription price from $250 to $100 a month around the same announcement, alongside the tier's storage and usage limits. [1][2][3]