What are Gemini Omni and the always-on agents Google showed at I/O 2026?

On 19 May 2026, at its I/O developer conference, Google announced Gemini Omni, a model that makes a video from any mix of text, images, audio and video. It also announced two kinds of always-on agents that keep working in the background: Gemini Spark, a personal agent in the Gemini app, and information agents in Search. According to Google's list of I/O announcements, Omni went live that day for paying subscribers, Spark went to trusted testers first, and information agents were due in summer 2026.

What is Gemini Omni?

Google describes Omni as a model that can "create anything from any input", starting with video (announcement by Koray Kavukcuoglu of Google DeepMind). The input can be any combination of text, images, video and audio. The output is video for now; Google plans other output types, such as images, later. The first model in the family is Gemini Omni Flash. Google does not describe the architecture. It only says that Omni combines Gemini's intelligence with its generative media models.

According to Google, three things set Omni apart:

• You edit by conversation. Each instruction builds on the previous one, and characters, physics and the scene stay consistent. In Google's example, a violinist is moved into the setting of a reference image, the violin is made invisible and the camera moves over the player's shoulder.
• You can reference anything: an image, a text, a video or an audio clip becomes one coherent clip. At launch, audio references are limited to voice.
• Google claims a better grasp of gravity, kinetic energy and fluid dynamics, and Omni draws on Gemini's knowledge, for example to make explainer videos.

Every Omni video carries SynthID, Google's invisible watermark, which you can check in the Gemini app, Gemini in Chrome and Google Search.

Omni Flash rolled out on 19 May 2026 to all Google AI Plus, Pro and Ultra subscribers worldwide, in the Gemini app and Google Flow, and at no cost in YouTube Shorts and the YouTube Create app. Developers and enterprise customers were promised API access in the following weeks.

What is Gemini Spark?

Spark is a personal AI agent in the Gemini app: software that carries out tasks on its own and keeps running around the clock. According to Google's Gemini app announcement, Spark runs on Gemini 3.5 and uses the Antigravity harness, the software layer around the model that plans work and calls tools. In the keynote, Sundar Pichai said Spark runs on dedicated virtual machines on Google Cloud, so it keeps working when your laptop is closed or your phone is locked.

Spark connects to Gmail, Docs, Slides and other Workspace apps. Google gave three kinds of use:

• recurring tasks or triggers, such as reading monthly credit card statements and flagging new or hidden subscription fees
• skills you teach it, such as pulling deadlines from your children's school emails into a daily digest for you and your partner
• complete workflows, such as turning meeting notes from emails and chats into a Google Doc and drafting the email that starts the project

Other services connect through MCP (Model Context Protocol), the open standard for linking AI tools to outside systems. Connections to Canva, OpenTable and Instacart launched at I/O, and Spark was to use them in the weeks after. The plans for summer 2026 also listed texting and emailing Spark, custom sub-agents and operating your local browser. Pichai added that Spark would work inside Chrome.

Google sets two limits. You choose whether Spark is on and which apps it connects to, and it is designed to ask first before high-stakes actions, such as spending money or sending an email. Google called Spark very early in its development and, citing safety, gave it to trusted testers first, in the week of I/O. A beta for Google AI Ultra subscribers in the US was planned for a week later.

What are information agents in Search?

Information agents are the first of what Google calls Search agents: agents you set up in Search that work in the background, 24/7 (Google's Search announcement). You describe what you want to follow. The agent watches blogs, news sites and social posts, plus Google's real-time data on finance, shopping and sports, and sends you a summarized update when something relevant changes, with the option to act on it. Google's example is apartment hunting: you list your requirements, and the agent tells you when a listing meets them. You can run several agents at once. Information agents were announced for summer 2026, first for Google AI Pro and Ultra subscribers.

The Gemini app got a smaller version of the idea on the same day. Once you opt in, Daily Brief works overnight and turns urgent Gmail messages and upcoming Calendar events into a morning digest with suggested next steps. It rolled out to Google AI Plus, Pro and Ultra subscribers, starting in the US.

What can developers build with it?

The harness behind Spark is open to developers as well. The Antigravity SDK gives programmatic access to the same agent harness, and you can host your agents on your own infrastructure. With Managed Agents in the Gemini API, in preview from 19 May 2026, one API call starts the Antigravity agent, built on Gemini 3.5 Flash, in an isolated Linux environment hosted by Google. There it can call tools, run code, manage files and browse the web. You define your own agents in Markdown files such as AGENTS.md and SKILL.md.

What are the risks of an agent in your inbox?

An agent that reads your mail and acts on it also reads any instructions an attacker hides in that mail. This is called indirect prompt injection. In June 2025, Google's own security team described this risk for Gemini in Workspace: instructions in emails, documents or calendar invites that tell the model to leak data or take other actions. Its defenses at the time came in layers, among them classifiers that detect such instructions, removal of suspicious links from answers, and a confirmation step before risky actions such as deleting a calendar event. Spark's promise to ask before spending money or sending an email is the same kind of confirmation step. The I/O announcements do not say how Spark handles instructions hidden in the content it reads.