Will your agent's work survive the cloud computer?
Coding agents now work on machines you never see. Before you leave one running overnight, know what it keeps when it pauses, resets or disappears, and make sure the work lives somewhere else too.
Published
In October 2026, users of one cloud agent service reported on X that their agents’ computers had been rolled back or reset without warning: files, SSH keys, skills and memory gone, and in one case a week of work. Whatever caused it, the lesson holds for every service: an agent’s work is only as safe as the place it’s stored.
What can be lost, from most fragile to least
- Running processes. The agent itself, its terminal, dev servers, background jobs. These stop whenever the machine pauses, sleeps, restarts or runs out of credit.
- Files on the machine’s disk. Your checkout, uncommitted changes, build caches, installed tools. These survive a pause on most services, but not a reset, a rollback to an old snapshot, or deletion.
- The machine’s setup. SSH keys, agent sign-ins, skills,
CLAUDE.mdandAGENTS.mdfiles, dotfiles. This is the part people forget, and the slowest to rebuild. - Work pushed somewhere else. Commits on GitHub survive anything that happens to the machine.
What each kind of machine keeps
| Running processes | Files on disk | The machine | |
|---|---|---|---|
| Claude Code cloud sessions | Stop when the VM pauses after a few idle minutes | Kept while paused | Can later be reclaimed; in-flight work is lost and the session restarts on a fresh VM |
| Docker Cloud Sandboxes | A session runs for up to 24 hours (1 hour by default) | Kept while paused | Paused sandboxes keep their state |
| Railway free agent VMs | Stop after the 60-minute window unless claimed | Deleted with the VM if it isn’t claimed within 24 hours | Kept once you claim it |
| boat.dev | Processes you started by hand don’t resume after boat stop | Kept in a snapshot | A stopped sandbox’s snapshot is kept for 14 days |
| A VPS you run | Stop on reboot unless you set up services | Kept | Yours, until you or the provider removes it |
| Cube | Keep running until you pause, or credit runs out | Kept on a persistent NVMe disk, also while paused | Yours until you choose Delete computer |
These are as of October 2026; the sources are at the end of the page.
Six habits that keep agent work safe anywhere
1. Make the agent commit and push as it goes
The single most useful rule. Put it in the instructions your agent reads:
## Saving work
- Work on a branch, never on main.
- Commit after each step that builds and passes its tests.
- Push the branch after every commit.
- Before a long task, write the plan to PLAN.md and commit it.
With this in CLAUDE.md or AGENTS.md, a reset costs you minutes, not a week.
2. Give every task its own worktree and branch
A worktree per task keeps parallel agents from overwriting each other, and a pushed branch per task means every agent’s progress is saved separately.
3. Keep the setup in a repository
Agent instructions, skills, shell config and a short setup.sh that installs your tools belong in a dotfiles repository. Then a fresh or replacement machine is ready in minutes. Keep secrets out of it; sign in again instead.
4. Write progress to files, not just to the conversation
Long tasks and loops should record where they are (a checklist in PLAN.md, a progress.json) and commit it. If the process stops, the next run reads the file and continues instead of starting over.
5. Back up what isn’t in Git
Databases, generated assets and anything in .gitignore won’t be on GitHub. Download an archive before any risky change, and always before deleting a machine.
6. Know what stops your agents
Find out what pauses your machine: idle timeouts, session limits, credit running out. A paused agent looks the same as a finished one until you open it. Check the machine’s status before trusting that a task is done.
How Cube handles it
Cube gives your agents their own Linux computer in the cloud with a persistent NVMe disk:
- Closing your laptop or the app never pauses it. Agents keep working in their terminals.
- Pausing keeps the disk. Running processes stop and your files stay, at $2 per vCPU a month. If credit runs out, the computer pauses the same way, and your files wait for you.
- The disk only goes away when you delete it. Delete computer removes the computer and its disk for good, so it’s the one step to take carefully.
- Backups are one click. In Machine, Download files saves your files, or everything on the persistent disk, to your Mac. See Back up your files.
Habits 1 to 4 still apply: Git is the backup that survives any computer, Cube included.
Questions
Does my agent's work survive if the cloud machine is paused?
On most services the files on disk survive and running processes don't. A paused agent stops mid-task, and its terminal sessions, background jobs and anything only in memory are gone. Commit and push often, so a stopped task can pick up from the last commit.
How do I back up work from a cloud coding agent?
Push it to a Git remote such as GitHub: that's the backup that survives any machine. Keep setup that isn't code, like agent instructions, skills and dotfiles, in a repository too, and download an archive of anything else before you delete a machine.
What happens to my files on Cube if my credit runs out?
Your computer pauses and its running processes stop, but its disk is kept. Add credit and resume it, and your files are where you left them. Only Delete computer removes the disk.
Sources
- Claude Code docs: Use Claude Code in the cloud
- Docker: Introducing Cloud Sandboxes
- Railway changelog #0309: free VMs without an account
- boat.dev docs: Stop a sandbox
- User report on X: a cloud computer reset six times in 11 hours (Oct 18, 2026)
- User report on X: a whole computer rolled back, a week of work deleted (Oct 16, 2026)
- Cube docs: Machine (pause, back up, delete)