Gemini 4 Argon as a background computer-use agent with Cua Driver
- Source
- x.com
- Date
1/ We had early access to Gemini 4 Argon. Its multimodal reasoning is a big step up: it reads dense screens, documents and spreadsheets the way a person does. With Cua Driver it works in your apps like a coworker next to you, without taking over your screen.
2/ Here it turns a whiteboard photo into a Keynote deck, working in the background while the person keeps typing in another app. 2,541 keystrokes and none lost.
3/ It reads a dense metrics dashboard and writes up the three biggest changes in Notes. Same setup: you keep working, it works alongside you.
4/ And where there's no API at all: it enters a month of expenses into a native ledger app, through the UI
Context
Cua posted a thread from its trycua account saying it had early access to Google's Gemini 4 Argon and paired it with Cua Driver so the model could work in a person's apps without taking over their screen. The split of responsibilities explains what that means. Google's computer-use documentation describes a loop in which the model looks at screenshots and returns UI actions such as clicks and key presses as function calls, while the developer has to implement the client that actually performs them. The same page lists gemini-4-argon among the models that support the tool.
Cua Driver is that client layer for native desktop software. Its repository describes MIT-licensed tools for inspecting and operating native apps and browsers on macOS, Windows and Linux, reachable through a CLI, or typed SDKs. Its background delivery lets an act without moving the user's pointer or taking focus, but only when the app and platform support it, so the thread's coworker-beside-you framing will not hold everywhere. The thread gives two ways to connect it: as an MCP server in Google Antigravity, where the thread notes Argon is on Ultra but any Gemini model works, or by passing Cua Driver's tools to gemini-4-argon as function declarations in a custom , with macOS setup shown for Cua Driver 0.35.
The evidence for how well Argon performs here is Cua's own demonstration. The thread describes three videos: Argon turning a whiteboard photo into a Keynote deck while the person types in another app, with 2,541 keystrokes and none lost; summarizing the three biggest changes on a metrics dashboard into Notes; and entering a month of expenses into a native ledger app that has no API. Those videos could not be retrieved for this explanation, and the thread offers no measured comparison with other models. For long agent runs the thread points users to a Cua Space, an isolated computer on hardware they own, which matches Google's own advice to run computer-use agents in a virtual machine or container rather than directly on the host.
- MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
Checking sign-in…
Loading comments…