Google Gemini News: Mac Agent Takes Over Your Apps
In the latest google gemini news, Google has given its Gemini assistant the ability to take over a Mac and do real work inside it. The company announced on Thursday that computer use is now rolling out in the Gemini app for Mac, turning the chatbot into an agent that can click, type, open menus and fill in forms across local files, desktop apps and the web. It is the most ambitious step yet in Google's push to make Gemini something that does things for people, rather than something that only answers questions. It arrives on the same day that Apple was reported to be building its own rival agent for the Mac.
The announcement, published on Google's official blog on November 19 and written by Michael Friedman, a group product manager for the Gemini app, frames the feature as a way to clear away repetitive work that normally takes dozens of clicks. Instead of jumping between apps and websites to finish a chore, a user describes the task in plain language and Gemini handles the clicking, typing, menus and forms. The work happens in an isolated background window, with a live preview shown inside the chat, so the user's own screen and cursor stay free.
For now, the rollout is deliberately narrow. Computer use is available in English in the United States, for personal Google accounts belonging to users aged 18 and over, on Google AI Plus, Pro or Ultra plans. It requires an Apple Silicon Mac running macOS Sequoia 15.0 or later, and Google says it is rolling out gradually, so not every eligible account will see it on the first day. The Gemini app for Mac is free to download from gemini.google/mac, but computer use sits behind the paid plans, which start at 4.99 dollars a month for AI Plus.
AI News: What Computer Use Can Actually Do on a Mac
The examples Google chose for the launch say a lot about who the feature is aimed at first. In creative work, Gemini can batch colour-grade photos in a favourite design application or set up a complex three-dimensional scene. In office work, it can pull weights, dimensions and addresses out of local PDF files and use them to fill in a multi-box shipping quote, or gather metrics and screenshots from a live dashboard and turn them into a presentation in Google Slides. For people selling things online, it can analyse photographs of clothing from a local folder, research prices, and stage listings on resale platforms. Each is a chain of small steps across more than one application, and that is the point: an agent's value is measured in the chains it can complete unaided.
To start a task, the user adds @Computer use to a prompt and describes what should be done. The app's connectors menu now includes a set of desktop apps, including Blender and Autodesk Fusion, the kind of professional tools where multi-step setup work eats up whole afternoons. Once the task is running, the user can watch each step in the live preview, pause the work, stop it, or intervene to confirm or block a particular action. Google says users choose which apps Gemini can access, and that key actions need approval before they happen.
The Mac app itself is still young. Google released the native macOS app on April 15, 2026, and on June 30 it brought its Gemini Spark agent to the Mac app, adding automation for local files and connections to other apps. Computer use builds on that foundation, extending the assistant from reading and organising files to operating the applications that hold them. Unlike older automation, which relied on purpose-built integrations or brittle scripts that broke whenever a button moved, a computer-use agent looks at the screen the way a person does and operates the button, menu or text field it finds. That generality is what makes the technology powerful and what makes it hard to get right.
Artificial Intelligence News: The Race to Build an Agent That Uses a Computer
Google is not first to this idea, and the history of the computer-use race explains why the launch matters now. Anthropic was the first to show the capability publicly, releasing computer use in public beta for its Claude 3.5 Sonnet model on October 22, 2024, aimed at developers rather than everyday consumers. OpenAI followed in January 2025 with Operator, a research preview for ChatGPT Pro subscribers in the United States, powered by a model called the Computer-Using Agent that combined vision with reinforcement learning. Operator ran in its own sandboxed browser, and in July 2025 OpenAI folded it into ChatGPT as part of a broader ChatGPT agent mode.
Google's own path ran through the browser. In December 2024, Google DeepMind unveiled Project Mariner, a research prototype built on Gemini 2.0 Flash that navigated websites through a Chrome extension. In October 2025, Google released its Gemini 2.5 Computer Use model in preview for developers, and in June 2026 computer use became a built-in tool in Gemini 3.5 Flash. The Mac launch is the moment that long research programme becomes a consumer product inside a mainstream app, with a subscription price and system requirements attached. The field has crowded quickly since. Meta launched its Muse agent in September, OpenAI followed with Dots, and Anthropic launched Claude Duet.
The testing behind these agents has become a story in its own right. Google's Chrome security team disclosed this week that it built an entire synthetic version of the web, called the Mirror Dimension, to test its browser-use agents safely. The environment holds thousands of simulated websites, built to look and behave like the real thing, where agents are tested against hidden prompt-injection attacks, the malicious instructions buried in web pages that try to trick an agent into taking unauthorised actions. The testing pipeline generates more than 1,000 attack test candidates per day for each security engineer, more than 30 times the number an engineer could create and test manually.
Apple News: Apple's Rival Agent and Tighter Rules for Everyone Else
The most striking part of the week's news may be the timing. On the same day Google launched computer use for the Mac, Bloomberg reported that Apple is building its own proactive AI agent, after spotting job listings on the company's website seeking a senior engineering manager to lead the work. The listing imagines an AI agent that can use a computer the way a capable person does, browsing the web, working across apps, running commands, writing and debugging code, and calling services, choosing the right tool for each step. Apple's team is already working on what the listing calls general-purpose computer-use agents, systems that plan and carry out multi-step tasks across apps, browsers, command lines, code and programming interfaces. Apple has eyed introducing the service as soon as the end of next year, according to Bloomberg, but development remains in the early stages. The company has not announced a product, and it did not respond to a request for comment from TechCrunch.
At the same time, Apple is making it harder for rival agents to gain full control of its computers. The forthcoming macOS 27.3, now in public beta, adds warning screens that prompt users to reconsider granting third-party agents permission to control their Macs. Bloomberg reported that the change raises the number of steps a ChatGPT user must take to enable third-party computer-use tools from four to fourteen. Apple said in October that it planned to introduce additional controls to ensure that users who genuinely wish to grant an app this extraordinary level of access can only do so with very explicit user action. Apple says the purpose is to protect users.
Tech News: A Week of Big AI Moves From Google
The Mac launch did not arrive alone. On Thursday, Blue Origin said Google had selected its New Glenn rocket to launch satellites for Project Suncatcher, Google's experiment in running AI systems in space. The launch is targeted for the first quarter of 2029 and will be Google's first heavy-lift flight dedicated to the project, carrying several computing satellites into sun-synchronous orbit. Each satellite will carry Google's tensor processing units, the company's in-house chips built to run AI programs, with computing capacity roughly equal to a rack of servers in a conventional data centre. A prototype satellite launched in October has already powered up its tensor processing unit hardware and generated a response using an onboard Gemma 4 model.
The same week brought a very different kind of pressure from inside the company. More than 500 Google employees and contractors signed a petition organised by the Alphabet Workers Union, titled Let Students Search, asking company leadership to allow schools to use Google Search without its AI features. The petition is addressed to chief executive Sundar Pichai, head of Search Elizabeth Reid, and education chief Chris Phillips. Google announced on November 3 that from early 2027, Gemini, Gemini Notebook, Gemini in Classroom and Search would be off until administrators turned them on, and districts that wanted AI off would lose Search too. The New York Times reported that school districts in New York City, Los Angeles and Chicago are considering switching to alternate search engines, and Senator Edward Markey wrote to Pichai on November 9 giving Google until November 30 to commit to letting schools disable AI features while keeping Search.
What Happens Next
For users, the immediate question is a practical one: does computer use on the Mac work well enough to trust with real tasks. Google's own framing is cautious. The feature is a rollout, not a finished product, and it is limited to a subset of users, in one country, in one language, on one kind of Mac. The controls Google describes, choosing which apps the agent can touch, watching each step, pausing or stopping a run, and approving key actions, are the right shape for a technology that is still earning trust.
For businesses, the enterprise angle is the one to watch. The tasks Google highlights, extracting data from PDFs to fill in shipping quotes, building presentations from live dashboards, and staging resale listings, are the kind of work that small businesses and creative studios pay people to do by hand. If an agent can do them reliably, the economics of that work change. Google has not yet said when computer use will reach work accounts. The company that can make an agent both capable and governable will win this race, and that will be judged not by announcements but by what happens the first time an agent, left alone in a background window, does something its user did not intend.
Source: Google