Cloud Agents: your computer is off,
the fix ships anyway.
Trigger a task from AgentsRoom on your desktop or your phone, and an AI coding agent runs it on a disposable machine in the cloud instead of your laptop. It clones the repository, works, commits, and pushes a branch with a written report.
No laptop in the bag, no remote desktop, no session to keep alive. The meeting point is git, which is exactly why the machine on your desk can stay switched off.
A cloud agent is a job, not a remote session. There is no screen to connect to and no terminal to keep open. A throwaway machine starts, clones your repository, runs Claude Code or Codex headless against the instruction you gave, commits what it changed, pushes a dedicated branch, writes a report, and then destroys itself.
The two machines never talk to each other. That single design decision is what makes the feature useful: since the rendezvous is a git branch and not a live connection, nothing on your side has to stay awake. Your laptop can be shut, flat, in a suitcase, or three thousand kilometres away.
What comes back is code and an account of what was done, never the agent's memory. So you do not resume a cloud agent, you review its work. Three actions are offered as equals: validate the branch, relaunch in the cloud with an extra instruction, or continue locally in a normal AgentsRoom session with the context already filled in. That last path is the one a purely cloud product cannot give you.
A user reports a bug. You are on a beach, with no computer.
This is not a thought experiment, it is the chain AgentsRoom was built to close. Every link already exists as a feature, and the last two are your own pipeline.
- 1
- 2
- 3
- 4
Minutes later, a branch is pushed with the fix committed
The run finishes and reports: the branch name, the commit, the files touched, and a written summary of the reasoning. You read it on the phone the way you would read a message.
- 5
Your CI picks the branch up and builds
This link is yours, not ours, and that is the point. A branch pushed to your repository is exactly the event your existing pipeline already listens for. Tests run, the build is produced, nothing was special-cased for AI.
- 6
Continuous delivery puts it live
A site is deployed, a mobile build is uploaded, a container is rolled out, whatever your delivery already does. The fix is in production and no computer of yours was ever switched on.
The lesson is about the chain, not the agent
Nothing here works because the model is clever. It works because every step was industrialised beforehand: a public place for reports, tickets scoped well enough to brief an agent that cannot ask questions, a repository whose branches trigger a pipeline, and delivery that does not need a human hand. AgentsRoom supplies the trigger and the scoping. Your CI/CD supplies the shipping. The phone is not an emergency IDE, it is the remote control of a chain that already runs without you.
Vacation is only the most obvious case
Every one of these is the same shape: the work has to happen, and the machine that usually does it is not available.
On call, on a Sunday
The alert fires while you are out. Reading the ticket on the phone and starting a run beats driving home to open a laptop, and the branch is waiting for review by the time you are back.
Your laptop died or got stolen
A dead battery, a spilled coffee, a bag taken at a station. Your projects are not hostage to one piece of hardware: the repository is the state, and a cloud machine can pick the work up from a borrowed phone.
A flight, a train, a commute
You can queue a run before boarding and read the report on landing. Ten hours of dead time turn into a branch waiting for you, without opening a laptop on a tray table.
Four tickets before going to sleep
Independent tickets do not need to be done one at a time. Send several to the cloud in parallel, each on its own clean machine and its own branch, and review the pile in the morning instead of watching them run.
The task is heavier than your machine
A long dependency upgrade or a wide refactor eats your laptop for an hour. Renting a machine by the minute keeps your fans quiet and your editor responsive while it happens somewhere else.
The office machine is not worth waking
A typo in production, a version bump, a copy fix. Not everything justifies remoting into your desk. A cloud run answers with a branch and a report, which is all a small change needs.
How a run actually goes
Five screens, from the moment you pick an agent to the branch waiting on your git host.
Pick where the agent runs
Every agent in AgentsRoom carries a Run on choice: this computer, another machine of your fleet, or the cloud. Nothing else about the agent changes, so the same role you already trust locally is the one that will run remotely.
Write the brief and set the run
The window takes the instruction, the agent (Claude or Codex on a cloud machine for now), the model from Haiku up to the flagship, and whether the work lands on a dedicated branch or the current one. Up to four files can ride along, such as a screenshot, a sketch or a log, copied onto the machine and destroyed with it.
Watch it, or do not
The run reports live: elapsed time, the model in use, the branch it is building on, and the instruction it was given. You can cancel it at any point. You can also close everything, because the machine does not depend on your window staying open.
The result is a branch on your git host
When it finishes, the work is where your tooling already looks for it: a pushed branch, a real commit, a diff you can review, and a pull request one click away. Nothing has to be imported back into AgentsRoom for the change to exist.
Then validate, relaunch, or take it back locally
The finished run gives you three moves of equal weight: keep the branch, send it back to the cloud with a follow-up instruction, or open it as an ordinary local agent session with the context already loaded. The third one is the bridge, and it is why a cloud run never becomes a dead end.
Four decisions that shape the feature
Most of what is surprising about cloud runs follows from these, so they are worth stating plainly.
A job, not a remote session
Live migration of a running agent session was examined and rejected as unfeasible. Treating a cloud run as a job instead removes the whole class of problems: nothing to reconnect, nothing to keep alive, nothing lost when your network drops.
Git is the only meeting point
The cloud machine and your machine never exchange anything but a branch. No forge is assumed either: credentials are held per git host, so this is not a GitHub-only feature.
A clean clone, every time
The agent works on a fresh copy on a throwaway machine, so it cannot touch your working tree, collide with a local agent, or leave anything behind. When the run ends, the machine and everything on it are gone.
It cannot ask you anything
There is nobody at the other end to answer, so the run either succeeds on the brief it was given or reports why it could not. This is the strongest argument for scoping a ticket properly: the quality of a cloud run is decided before it starts.
Your public backlog is the natural source of cloud runs
A cloud agent works alone and cannot ask a single question, so a scoped ticket makes a far better brief than a line written in a hurry. That is not a limitation to work around, it is a habit worth having anyway.
It also explains why the two features belong together. Your public backlog keeps collecting user reports while nobody is at the machine, overnight, over a weekend, over three weeks in August. Those tickets are exactly what a cloud run is for: already written down, already described by the person who hit the problem, and waiting for a decision rather than for a keyboard.
Metered in minutes of machine time, not in tokens
Every other metered part of AgentsRoom spends tokens. A cloud run rents a real machine billed by the second, so it is counted in minutes of machine time instead, and that difference drives the rules.
Included minutes come with the Plus and Pro plans and reset with the calendar month. Minutes you buy on top never expire, because confiscating paid credit at the end of a month would be indefensible. Every run is bounded twice, by a maximum duration and by how many can run at once, so a task that goes wrong cannot spend an unbounded amount of machine time.
The free tier includes zero cloud minutes, and that zero is a rule rather than a ceiling: a run is real hardware rented for real money, which is not something a free plan can absorb. Cloud Agents is currently in beta and opens through the self-serve beta programme.
What it does not do, stated plainly
A feature page that only lists strengths is a feature page you cannot trust.
- Claude and Codex are the only agents that run on a cloud machine for now. The other providers stay local.
- A run has a maximum duration and stops when it is reached. It is sized for a scoped task, not for an afternoon of open-ended exploration.
- The agent cannot ask a question mid-run, so an ambiguous brief comes back as an ambiguous branch.
- Runs are started by you or by a schedule, from the desktop, the phone or a backlog ticket. There is no webhook trigger, so a failing pipeline cannot start a run on its own yet.
- Nothing is merged for you. A run answers with a branch, and what happens next is your review and your pipeline.
FAQ
Do my agents still run locally?
Yes, and that is still the default. AgentsRoom runs your agents through the CLI you already pay for, on your own machine, against your real working tree. Cloud Agents is an explicit per-run choice, taken task by task, not a mode the product moves you into.
Can I really start a coding agent with my computer switched off?
Yes. The mobile app reads your projects and your backlog from the AgentsRoom API rather than through your desktop, and a cloud run happens on a rented machine, so nothing on your side needs to be awake. That is the difference with remote control of a local agent, which does need your computer to be on.
What comes back when a run finishes?
A branch pushed to your repository with real commits, plus a written report: what was changed, how many files, and the reasoning behind it. You do not get the agent's memory or a session to resume, because the run and its machine no longer exist.
How does this connect to my CI/CD?
Through nothing special, which is the point. A cloud run pushes an ordinary branch to your ordinary repository, so whatever your pipeline already does on a push happens exactly as before. Build, test, preview deployment, mobile build, continuous delivery: none of it needs to know an AI wrote the commit.
Is the code safe on a machine I do not own?
The machine is disposable and dedicated to one run. It clones the repository, does the work, pushes the branch and is destroyed, taking the clone and any attached files with it. Git credentials are held per host and never handed to the agent as text.
What does a cloud run cost?
It is counted in minutes of machine time, not in tokens, because the run rents real hardware billed by the second. The Plus and Pro plans include monthly minutes, extra minutes can be bought and never expire, and the free tier includes none. Current amounts are on the pricing page.
What makes a good brief for a cloud agent?
A ticket somebody already took the trouble to describe. The agent works on a clean clone and cannot ask you anything, so it needs the acceptance criteria, the file or area involved, and how to tell that it worked. A well scoped backlog ticket is usually a better brief than a sentence typed in a hurry from a phone.
Can I take a cloud run back into a normal session?
Yes, and it is one of the three actions offered when a run ends, alongside validating the branch and relaunching in the cloud. Continue locally opens a regular AgentsRoom agent with the context pre-filled, so exploratory follow-up work happens where you can talk to the agent.
How is this different from Warp Oz or Devin?
The execution model is close, the position is not. Cloud Agents is one mode inside a product whose default is local execution on the CLI subscription you already hold, so a cloud run can be handed back to a local session with its context. Tools built cloud-first have no equivalent of that return path.
Is Cloud Agents available to everyone?
Not yet. It is in beta and opens through the self-serve beta programme, which asks for an active Pro subscription, an account at least three months old, and a track record of feedback the team actually acted on. The Beta tab in Settings shows which conditions you already meet.
Features that make this workflow work
Public Backlog
The public board where users file the reports that become cloud runs, with a reporter thread attached to each one.
Mobile Companion
The iOS and Android app that lets you read the backlog and start work from a phone, with no computer of yours awake.
Ticket Scoping
Turn a vague report into a brief precise enough for an agent that cannot ask you a single question.
Scheduled Tasks
Fire recurring agent runs on a schedule, the other way to make work happen when nobody is at the keyboard.
Remote Fleet
Run agents on every machine you own from one window, when the machine you want is yours rather than rented.
Remote Control
Pilot the agents running on your desktop from your phone, for when your computer is on and you are not near it.
Further reading
Working on Vacation With AI Agents (Without Your Family Noticing)
Close the shop for three weeks, or be the person on the laptop at the beach. AI coding agents make a third option real: the setup that keeps client projects moving on ten minutes a day.
Background Coding Agents: Putting Your AI on the Night Shift
A coding agent doesn't need you watching it. Here's how to run agents in the background while you work on something else, and how to let a whole fleet code overnight while you sleep.
Driving Your Coding Agents From Your Phone While They Work
Your agents code for 20 minutes straight. You don't need to sit in front of the screen. Here's how to track and unblock them from your phone.
Stop being the single point of failure
Set the chain up once: a public backlog, tickets worth briefing an agent with, and a pipeline that ships on a branch. After that, a fix can leave your hands without you opening a laptop.
Companion app: monitor your agents on the go
Bring your own: Claude, Codex, Antigravity CLI, or other AI provider.
Push bugs and requests straight to your public backlog.
A glimpse of AgentsRoom in action.