Explainer
An AI agent works across devices when its conversation, its memory and its running task live on a deployment rather than on any device you hold: continuity is a property of where the state lives, not of a mobile app. Most people asking this have an agent bound to one machine or one browser tab, and want to start a task at a desk and finish it on a train. On Olano the deployment owns all three, so every device is a window onto the same agent - the dashboard on a laptop, the same dashboard in a phone browser, WhatsApp, Telegram. This page explains what that means in practice, and the one thing it does not mean.
Ask where the state lives and the answer usually takes one of three shapes. An agent running as a process on your own machine keeps everything inside that process: the conversation, whatever it has learned, the task it is halfway through. Move to another device and there is nothing there to move to. A browser-based tool often holds the running task inside the page that started it, so the history may be saved while the work belongs to that tab. And a consumer assistant with an account syncs the chat history between phone and laptop, which covers one of the three things: the transcript moves, and whether the memory and a task in progress move with it is up to that product.
None of this is about an app. It is one fact in three costumes: the state lives in a session, and a session belongs to one device.
Olano keeps the three things that need to be in one place on the deployment itself, not on anything you hold, and the device question then answers itself. There is nothing to sync, because none of it is on the device. A laptop, a phone browser, WhatsApp and Telegram are four windows onto one agent that is already holding all of it.
Each would be enough to tie an agent to a device if it lived anywhere else.
Stored on the deployment and listed on every surface. It survives switching the model mid-conversation and a restart of the deployment, and a copied link opens exactly that thread in any browser you are signed in on. Opening an agent on a new device lands you in its latest conversation, whichever surface that came from, not in an empty one.
Markdown files in the agent's workspace - MEMORY.md for durable facts and a dated daily log - loaded into every conversation the agent has, on any channel. Self-learning distils recent conversations into them in the background, and the agent can write to them itself during a conversation. See memory and knowledge.
A turn is a job the deployment runs. Your screen attaches to it as a viewer, and a viewer arriving or leaving changes nothing about the job.
Same agent, same memory, same running task. What differs is what each window can show.
| Window | What you see | Worth knowing |
|---|---|---|
| Dashboard on a laptop | The full chat: the reply forming, tool calls as they run, files the agent produced as cards with a Download button, and every conversation the agent has had on any surface. An agent's row shows a pulsing dot while it is working and a count of replies that landed while you were elsewhere. | Each conversation carries a badge naming where it came from - TELEGRAM, WHATSAPP and so on - so a thread that arrived from a messenger reads differently from one typed here. |
| Dashboard in a phone browser | The same dashboard at the same address. It can be added to the home screen, where it behaves like an app: its own icon, full screen, sign-in kept between launches. On Android it then appears in the share sheet, so a link or a note shared from another app opens the chat with that text already in the message box, waiting for you to send it. | Nothing to download; on an iPhone the home-screen install is Safari's Share menu. Agents run on the deployment, so sending a message needs a connection; an installed copy only keeps the last screen readable when the network drops. |
| WhatsApp or Telegram | An ordinary chat with the agent. It answers as itself, with the same instructions, tools and memory as on any other channel, and a file it produces arrives as an attachment. Set up per agent under channels and connectors. | The conversation appears in the dashboard list as well, so a colleague can open it from a laptop and see who said what. A reply typed there is recorded and badged in the thread, not sent out to the messenger. |
| A colleague's screen | Everyone who can chat with an agent shares its whole conversation list and can pick up any thread in it - a scheduled run, or a chat someone else started. | Messages show who wrote them, and which agents each person can see is set per person. See multi-user agents. |
When you send a message, the deployment starts a turn and your screen attaches to it as a viewer. Close the tab, lock the phone, switch apps, walk from Wi-Fi onto mobile data: the viewer goes away and the turn does not notice. The agent keeps reading, calling tools and writing, because none of that was happening in your browser.
Come back - the same tab, a new one, or another device - and the screen rejoins the reply exactly where you stopped seeing it. It fills in the words, the tool steps and the files you missed, then carries on live. A reply that finished while you were away is simply there. A conversation the agent is still working on shows as busy in the list on every device, including one that never saw the turn start, and the working indicator names the surface the turn is running from, so a laptop can watch a reply that began on Telegram.
Two consequences follow. A conversation runs one reply at a time, so sending again from your phone while the laptop's turn is still running does not start a second one; the message either steers the reply in progress or waits its turn. And two status lines mean different things. Reconnecting means the reply is catching up and nothing was lost, so there is nothing to resend. Connection lost means what it says - the reply is genuinely gone, because the deployment restarted or the reply ended long enough ago to have been cleared - and that is when a Retry button appears.
Because closing a window ends nothing, Stop is the control that does. It tells the deployment to end the turn, which is why it works from any device: you can stop from your phone a task you started on a laptop that is now shut in a bag, whether or not the tab that started it is still open. Closing the tab is not a stop.
There is no native mobile app, and we would rather say so than imply one. It is the dashboard in a phone browser, installable to the home screen, plus the messaging channels - and a messenger is already the app you have open.
Three limits, before you lean on any of this. A conversation that arrived over WhatsApp or Telegram can be opened and continued from the dashboard, but a message you type there is recorded in the thread and badged, not sent out to the messenger. A chat typed in the dashboard and a chat over a messenger are separate threads that share one memory, rather than one thread that follows you between apps: what was said in one reaches the other only through the memory files - when the agent saves it there during the conversation, or when the next background self-learning pass distils it, which runs on a schedule rather than after every message - and never as a shared transcript. And the deployment has to be running; a machine you carry cannot be the place the state lives, which local vs cloud AI agents covers in full.
Yes. The agent, its conversations and its memory live on the deployment, not on either device, so the dashboard on a laptop and the same dashboard in a phone browser are two views of one agent. Open the same conversation on both and each follows it as it happens, including a reply that is still being written.
On Olano, yes. A turn runs on the deployment and your screen is only a viewer, so locking the phone, switching apps or losing signal removes the view and leaves the work running. When you come back, the screen rejoins the reply where you stopped seeing it, fills in the tool steps and files you missed, and carries on live.
Nothing needs syncing, because the conversations are not stored on a device. Every conversation is held on the deployment and listed on every surface, with a badge showing where it came from - a Telegram thread, a scheduled run, a chat a colleague typed. A device that was asleep or offline catches up on its own when the tab comes back.
There is no native app. The dashboard works in a phone browser and can be added to the home screen, where it behaves like an app with its own icon, full screen and a sign-in kept between launches; there is nothing to download. For many people the practical mobile surface is WhatsApp or Telegram, where the agent answers with the same instructions, tools and memory as anywhere else.
It shares the same memory: durable facts in MEMORY.md and the recent daily logs are loaded into every conversation on every channel. The WhatsApp chat and the dashboard chat are separate threads, though, so something said in one reaches the other only once it is in those memory files - saved by the agent during the conversation, or distilled by the background self-learning pass, which runs on a schedule rather than after every message - not as a shared transcript.
Press Stop. It tells the deployment to end the turn, so it works from any device and whether or not the tab that started the reply is still open. Closing a tab, locking a phone or losing signal only removes your view of the run; the run itself continues until it finishes or you stop it.
Where to go next.
The buyer's checklist; this page is the long answer to one of its nine questions.
The memory every window onto the agent shares, and how it is built.
Setting up WhatsApp, Telegram and the other channels, per agent.
Why a machine you carry cannot be the place the state lives.
Tell us which task you keep losing between devices. We will show you the same thread open on a laptop and a phone at once, and what happens when the phone locks halfway through.
