Google's most revealing Gemini announcement is not the video model, even though video is the part with the easiest demos. It is the small, almost administrative phrase attached to Gemini Spark: runs 24/7. That sounds like infrastructure copy until you sit with it for a minute. A chatbot answers while you are there. A background agent works while you are not.

The difference is not cosmetic. In Sundar Pichai's I/O note, Spark is described as powered by Gemini 3.5 and Antigravity, built for long-horizon tasks that can keep going in the background. The new Gemini 3.5 Flash is the model underneath that pitch, with Google calling it available across the Gemini app, AI Mode in Search, Antigravity, the Gemini API, AI Studio, Android Studio, and enterprise products. This is not a lab release. It is a distribution event with a model announcement folded inside it.

I wrote earlier this month about Google turning I/O into a Gemini argument, and the argument has hardened since then. Gemini is no longer being presented as one surface. It is the layer that Google wants to move through Search, Android, developer tools, enterprise software, creative tools, and whatever form of assistant hardware survives contact with real life. Spark makes that plain because it changes the unit of interaction from a prompt to an errand.

That is where I get wary. Not because background agents are useless. Quite the opposite: they are useful in the exact boring places that make software stick. Monitor this. Reconcile that. Keep trying until the slot opens. Pull the materials together before I come back. The dull examples are the serious ones, because they don't need wonder. They only need enough reliability that a person stops watching.

Google's advantage is that it owns so many places where not watching is already normal. Search waits. Gmail waits. Drive waits. Android waits in your pocket all day, and Chrome sits between intention and almost everything else. OpenAI can make a cleaner interface; Anthropic can make a better argument about restraint. Google can place the agent inside the room where the task was already going to happen.

The danger is also Google-shaped. A background agent turns permission into a standing condition. It asks for less attention at the exact moment it needs more trust. If Spark is checking, filing, drafting, booking, comparing, or nudging on your behalf, then the old question of "what did the model say?" becomes less useful than "what has it been doing?" That is a different audit problem, and a different kind of intimacy with the machine.

Gemini Omni still matters. Google says Omni can take images, audio, video, and text as input and create video first, with Gemini Omni Flash rolling into the Gemini app, Flow, and YouTube Shorts. I covered the creative side of that in Video Becomes the Prompt, where the edit starts to live inside conversation. Spark is the plainer version of the same strategic move: move the work out of a specialist tool and into the ambient Google layer.

The phrase "agentic Gemini era" is ugly, but the ugliness is useful. It has the bluntness of an internal roadmap escaping into public. Google is not asking whether people want a more charming chatbot. It is asking whether the next interface can be a thing that stays awake after the tab is closed. I don't think users have really consented to that idea yet. I also don't think consent will arrive as one grand yes or no. It will arrive as a thousand small defaults, each one too convenient to refuse on its own.

Sources: