Google could let Gemini punt tasks to Spark without even opening the app

Google announced Spark, its autonomous AI agent that performs a sequence of tasks without intervention, at Google I/O 2026. Since then, Spark has rolled out in beta to Gemini (so far only to paying Google AI subscribers) on the web and mobile versions. On top of that, Google can be seen making efforts to make Spark’s skills easier to manage from mobile, especially as it looks to sunset Gems later this year.

In line with those efforts, we’re now seeing evidence that Google is making it easier to assign tasks to Spark, even mid-conversation.

Last month, we learned Google is working on an option to offload tasks to Spark directly from a regular Gemini chat instead of switching over from the sidebar first. While the feature hasn’t rolled out, we’re seeing new evidence suggesting a similar action to engage the agent could be executed using the floating Gemini overlay, even without opening Gemini fully. These hints were spotted as part of version 17.56.14 of the Google app, which we tweaked to enable a preview, even though the feature isn’t live for everyone yet.

Users would likely be able to access the Assign option by tapping the plus (+) button or expanding the overlay using the grab handle. Here, they would likely see an option to “put Gemini to work” on a specific task. Currently, we see a toggle here instead of a regular button. Once this toggle is enabled, the text on the overlay changes from “Ask Gemini” to “Describe a task.”

This is especially favorable for people who tend to launch Gemini by saying the “Hey Gemini” or “Hey Google” wake words or long-pressing the power button.

From what it seems, you cannot choose to engage Spark by default for every query on Gemini, and you will be required to enable the toggle manually for each chat.

Along with the ability to “Assign” tasks for Spark to run in the background, Google also appears to be adding the option to select the model directly from the Gemini overlay. Currently, the option to select the model is limited to the main chat interface. In its current form, the option opens the list of supported models but doesn’t change them, so it’s hard to say whether that is the exact functionality Google has planned. We expect to see more progress on this front in the coming months.