I spent a few days testing Gemini’s new task automation on a Pixel 10 Pro and a Galaxy S26 Ultra. For the first time the assistant can actually operate apps on your phone for you. Right now it is limited to a small group of food delivery and rideshare services and it is still in beta. It is often slow and clumsy, and it does not solve major phone problems yet. Still, seeing an assistant do real app work on a real phone is impressive and feels like an early look at what comes next.

How it behaves

By default Gemini runs automations in the background. You can watch progress if you want, and when you do the phone shows short status lines at the bottom of the screen describing each step. That visual feedback is helpful. In one test I told Gemini to order a chicken combo plate. The menu showed half portions, so Gemini added two half portions to equal a full order. That kind of on-the-fly reasoning is neat to watch.

At other times Gemini gets stuck on things that a human would solve in a second. It once had trouble finding a side that was visible at the top of the menu, and the whole attempt to build my dinner order took about nine minutes. That is not fast. If you need a ride right now, you should still tap the app yourself.

What it does well

  • Background work: It can continue an automation while you use other apps or even while the phone is idle.
  • Stops before final confirm: Gemini completes steps up to the final confirmation so you can check the order or ride details before you commit. In my tests it never finished an order without my explicit approval.
  • Access to email and calendar: Because it can read calendar and email data, it can infer things like flight times. I gave it a calendar event for a flight and asked it to schedule a ride. After a little guidance it suggested departure times and then reserved a ride in about three minutes with no more input from me.
  • Accuracy: Most of the time it gets the order right and needs only minor tweaks to the final screen.

Where it fails

  • Slow: Many tasks take several minutes. Some steps involve hesitation and repeated taps.
  • Fragile with dialogs and permissions: Failures tend to occur early when an app needs a permission or the delivery address is wrong. Those require a human to fix things, then the automation can be restarted.
  • Poor explanation: When it cannot complete a step it often does not tell you why, so you must inspect the app to find the issue.

Why this matters

Watching an AI scroll and tap through apps reveals an obvious truth: current apps are built for humans, not for automated agents. A human-focused user interface includes ads, promotional banners, and visual cues that look great to people but add noise for an assistant. If designers built app interfaces with machines in mind, the assistant could be faster and more reliable. Industry work on structured interfaces and protocols aims to address this. In the meantime Gemini uses reasoning to navigate human-style apps when a structured option is not available.

Google's head of Android has said that when app-level integration methods are not present, Gemini falls back to reasoning through the app. That approach works, but it is a stopgap until more apps adopt developer-friendly methods such as protocol-based integrations or platform app functions.

Bottom line

This version of task automation is awkward and slow, but also clearly capable. It rarely completes orders without your confirmation, it can pull flight and calendar details to plan rides, and it handles some menu logic intelligently. It is not ready to replace you for time-sensitive tasks, but it is a meaningful first step toward assistants that actually use apps for you. In other words, imperfect now, promising later.

Photography credit withheld in this rewrite.