Google is turning its Gemini desktop app into two apps in one. A new switch lets you flip between Ask, the familiar chat box, and Assign, a mode where the AI picks up a task and finishes it inside a folder you choose, largely on its own.
This is not a small tweak. Assign is built on Gemini Spark, the agent Google introduced earlier this year that runs continuously and connects to Workspace tools like Gmail and Docs. In testing, Assign can also link to a second computer running Gemini, and it can turn on a skill called computer use, which lets the AI see your screen and click through apps the way a person would, rather than just answering questions in a text box.
None of this is unique to Google. OpenAI already ships a desktop version of Codex that splits work the same way, between chatting and letting an agent run a task in the background. Anthropic's Claude can already open apps on your computer, browse the web, and fill in spreadsheets on command. Google is playing catch up here, not leading, and the fact that all three companies are converging on the same design, a chat mode and a work mode, tells you this is where the whole industry is headed, not just one company's idea.
For a normal business, the shift that matters is not the toggle itself. It is what the toggle represents. Until now, AI tools mostly waited for you to type a question and gave you an answer back. An agent in Assign mode does the opposite: you give it a folder and a goal, and it goes and does the work, opening files, running steps, and only checking back in when it needs a decision from you. That is a real change in what these tools can be trusted to do without someone watching every step.
That trust question is exactly where the risk sits. Any tool that can click through your apps and touch your files needs clear boundaries on what it is allowed to see and change. Industry researchers already track a rising number of security incidents tied to AI agents that were given more access than anyone meant to grant, simply because nobody set the boundary in advance. A folder-scoped agent is a reasonable first step toward limiting that exposure, but it only works if whoever turns it on actually thinks about which folder that should be.
Right now, none of this is available to the public. Google has these features in closed testing with no announced release date, and features like the Obsidian connection and the plugin tab are still rough. But the pattern across Google, OpenAI, and Anthropic is consistent enough to plan around: within the next product cycle, mainstream AI apps will come with an agent mode built in as a standard feature, not a novelty. Operations and IT leaders have a real window right now to decide, before it lands on everyone's desktop, which files and apps they would actually be comfortable letting an AI operate without a human at the keyboard.