Cloud agents
Agents that run each task on a machine in someone else’s cloud, take work from a chat, ticket or web app, and come back with a pull request.
How they work
A cloud agent gets its own computer per task: a VM or container with your repository cloned, its dependencies installed and, often, a browser. You start it from a web app, Slack, an issue tracker or the CLI, and it keeps working after you close your laptop. Devin, Codex cloud and Claude Code’s cloud sessions all work this way; OpenAI’s first Codex cloud release said tasks typically took 1 to 30 minutes and ran with internet access disabled.
The separate machine is both the point and the cost. It can’t damage your laptop, and you can run many tasks in parallel, but its environment has to be set up to match yours (setup scripts, secrets, network allowlists), and your code runs on the vendor’s infrastructure. Cognition, which makes Devin, frames the split this way: a local agent makes you faster, a cloud agent does the work while you’re not there. Several products now let you start a session locally and hand it to the cloud.
Technologies
What all ai coding agents have in common →Sources
- Cognition: Devin in Windsurf, on local and cloud agents (Apr 2026)
- Cognition: Devin CLI, start local and hand off to the cloud (Apr 2026)
- OpenAI: Introducing Codex (cloud sandbox per task, 1 to 30 minutes, internet disabled; May 2025) · archived copy, original
- Codex docs: Codex Cloud (each task has its own workspace)
- Claude Code docs: Use Claude Code in the cloud (isolated VMs, network limits)