--- title: "Android XR and on-device agents: a change, not a gadget" url: "https://vucale.com/blog/android-xr-and-on-device-agents" description: "Agents that live in glasses and phones arrive as a platform decision. What to settle first: dependence on one vendor, rules for agents, trust, and a test without hardware." --- # Android XR and on-device agents: a change, not a gadget October 13, 2026·4 min read·Editorial team Three young colleagues around a light wooden table, one holding plain unbranded glasses next to a phone, the others sorting pastel sticky notes into two columns An operations director comes back from a conference convinced that the next interface for field staff is a pair of glasses. IT replies that the company runs on another vendor’s phones and has no device management for anything with a camera on someone’s face. HR asks who is going to explain to the workforce that an assistant will now see what they see. Within a week the conversation has stopped being about glasses. That is the right instinct. Android XR is Google’s platform for XR headsets, display glasses and audio glasses, with Gemini as the assistant layer, developed with Samsung and Qualcomm. The first device was Samsung’s Galaxy XR headset; XREAL Aura is the first priced display-glasses device; Samsung has announced its own Gemini glasses for November 2026. At Google I/O in May the company showed audio glasses that work with a phone, Android or iOS, where the agent carries out multi-step tasks on the phone in the background and the glasses act as its voice and camera. Put plainly: the agent no longer lives in a browser tab. It lives on a device the employee carries, it can act, and the platform underneath it belongs to one vendor. That is an organisational change, and it deserves to be treated as one. ## Where companies get it wrong **The device is chosen before the platform.** A pilot buys whichever glasses are available, and by the time it ends the company has quietly picked an operating system, an app store, an assistant and an identity provider. None of those choices was made on purpose. **Agent permissions are treated like camera permissions.** The familiar question is whether an app may use the camera. The new one is what the agent may do on the user’s behalf: open an app, send a message, fill in a form, place an order. Those are actions, and they need rules the way a new employee’s access does. **Where the work runs is taken on trust.** We have found no public breakdown, from Google or from XREAL, of how much of the agent’s work happens on the device, on the phone or in the cloud. That may differ by feature and change with updates. Until you have it in writing, treat everything the camera and microphone capture as potentially leaving the building. **People are told, not asked.** A wearer who feels watched by their own assistant, or colleagues who feel filmed, will find ways not to use it. Trust here is not a communication task at the end; it is a design constraint at the start. **Lock-in is discovered at renewal.** Apps built for one assistant, data in one vendor’s cloud, processes rewritten around one device: each is reasonable on its own, and together they decide your next five years. ## What to settle instead **Decide the platform question as a platform question.** Who owns the identity, the data and the app distribution? What does leaving look like: which parts of your work could move to another vendor’s glasses, and which would have to be rebuilt? Write the exit down before the entry. We make the same argument in [testing a platform on your own data](https://vucale.com/blog/test-the-platform-on-your-data), and it holds for devices. **Write rules for agents, not just for apps.** List the actions an agent may take alone, the ones it may prepare for a person to confirm, and the ones it may never take. Keep a record of what it did. A person stays accountable for the outcome, as with any other tool. **Start with the people who will wear it.** Ask them where the assistant would help and where it would intrude. Agree on visible signals, such as the privacy light, and on places where devices come off. The rules on unmanaged AI tools from [our piece on shadow AI](https://vucale.com/blog/your-people-already-use-ai) apply here too: an approved route beats a ban. **Keep your own logic outside the device.** Process knowledge, prompts and integrations that live in your systems can be pointed at another platform later. Logic built into one vendor’s assistant cannot. ## A test you can run without buying hardware You do not need glasses to learn most of what a pilot would teach. The [Android XR SDK](https://developer.android.com/xr) includes a glasses emulator, and for glasses apps it offers Jetpack Projected: the app runs on a host device, such as a phone, and projects the experience to the glasses. Jetpack Compose Glimmer covers interfaces for see-through displays, and a Device Availability API tells an app whether the glasses are actually being worn. That is enough for a two-week exercise. Take one task, build or mock up the flow on the emulator with a phone as host, and run it past the people who do the work. Note which steps the agent would take on its own, which data it would need, and where it would send it. If you cannot answer those three on paper, the hardware will not answer them for you. The device side of the same decision is in the companion piece on [whether to pilot XR glasses with an AI agent](https://vucale.com/blog/should-you-pilot-xr-glasses-with-an-ai-agent). We sell no devices and no platforms; as the [homepage](https://vucale.com/) says, our work is diagnosis, a working pilot and a change that takes hold. ## The question for this week Ask whoever owns IT and whoever owns data protection one question each: if an AI agent on a device took an action in a customer’s system tomorrow, how would we know, and who would answer for it? If neither can answer, the first project is not a pilot. It is the rules. - Android XR - AI agents - vendor lock-in - change management ## Keep reading - [Agent factories will not stay in IT](https://vucale.com/blog/agent-factories-will-not-stay-in-it) - [The pilot worked. Scaling is a different project](https://vucale.com/blog/the-pilot-worked-scaling-is-a-different-project)