You set Claude Code to work on a UI component, step away for 20 minutes, and come back to a dropped connection. No log. No way to resume. Whatever the AI built before losing the link is gone.
That experience is common enough with remote AI coding tools right now that it's worth examining directly - not as a bug report, but as a design failure.
The Session Drop Problem
Remote AI coding works like this: instead of the AI running code on your local machine, it operates in a separate environment (cloud or otherwise). You give it a task, it works independently, and ideally returns a result without you watching every step.
The appeal is real. Long-running tasks, automated pipelines, environments where you don't want AI tools touching your local filesystem - remote execution makes sense in those cases.
But the current reality is sessions that drop with no recovery path. The connection goes, the work disappears, and you start over. There's no state log showing where it was when it disconnected. There's no task queue that resumes where it left off. Just a blank screen and a lost hour.
Reports of Claude Code's remote session going completely unavailable for hours at a stretch aren't edge cases. They're the kind of friction that makes a tool unsuitable for anything you can't afford to redo.
Building UI While Blind
The second problem is more fundamental for anyone doing frontend work. When you ask a remote AI coding tool to build a UI component, you get back code and maybe a description. You do not get a rendered preview.
You're reviewing a diff for a button you have never seen render. You're approving layout changes without seeing the layout. Accepting a change because the description sounds right is not reviewing - it's guessing.
Local tools like Cursor and Cody sidestep this because they run inside your actual development environment. Open the browser, hit refresh, see the result. Remote tools don't have that path by design, and none have built a working substitute yet.
What Would Actually Fix This
Session persistence when connections drop is table stakes. Any serious distributed system logs state. Remote AI coding tools should too. If a connection drops 18 minutes into a 20-minute task, resuming from minute 18 is not a stretch ask - it's a baseline.
Visual previews are harder but not impossible. A screenshot of the rendered component returned alongside the code would close most of the gap. Some tools are experimenting with browser-in-the-cloud approaches. None have made it reliable enough to count on in real workflows.
Remote AI coding has a legitimate future. But right now it asks you to trust a black box that goes offline regularly and cannot show its work.