The GenAI Field Guide · Trending
Coding agents that run on the vendor's virtual machines instead of your laptop, so a task keeps going after you close the lid and you can check on it from a phone.
On 23 September 2026 Anthropic took Claude Code cloud sessions out of research preview, and on 29 September OpenAI used DevDay to make Codex cloud environments persistent and shareable across desktop, web and mobile. Both move the agent loop off your machine into an isolated VM that clones your GitHub repository, works on a branch and hands back a diff or a pull request. If your team already uses a coding agent, this changes what you can delegate: long, well-specified tasks run in parallel and unattended, and the review step becomes the bottleneck.
A cloud session is the same coding agent you run in a terminal, but the process, the shell and the checkout live on a virtual machine the vendor manages. Anthropic's docs define it as a Claude Code session that runs on cloud infrastructure instead of on your machine, which keeps running after you close your laptop and can be checked or steered from any device. You start one from the browser at claude.ai/code, the Code tab in the Claude mobile app, the Desktop app with Cloud selected, or the terminal with claude --cloud. Scheduled and triggered routines also run as cloud sessions.
OpenAI's version is Codex cloud, part of ChatGPT. The DevDay change is that a cloud environment, meaning the repositories, dependencies, tools and access settings a task needs, is now something you prepare once, publish and reuse. Each new task gets its own isolated workspace from that published environment, and in Enterprise workspaces an environment can be shared so colleagues start from the same base. OpenAI's DevDay notes say cloud work continues across web, mobile or desktop while your own machine is offline.
Under the marketing, neither of these is a new model or a new kind of agent. It is a change in where the agent runs and who owns the sandbox. That matters for three practical reasons: the work no longer dies when your laptop sleeps, you can fan out several tasks at once without juggling local worktrees, and the security boundary (network egress, credentials, what the agent can push) is now configured on the vendor's side rather than inherited from your workstation.
For Claude Code, each session gets a fresh VM running Ubuntu 24.04 on x86_64 with 4 vCPUs, 16 GB of RAM and 30 GB of disk, per Anthropic's docs. The repository is cloned from GitHub, an optional setup script runs as root, network access is set from the environment's level, and Claude works until it reaches a stopping point and pushes a branch. You then review the diff, leave inline comments, open a pull request, or keep the conversation going. A setup script that finishes in roughly five minutes is snapshotted, so later sessions start with dependencies and Docker images already on disk; the cache rebuilds after about seven days or when the script or allowed hosts change.
Isolation is the heart of it. Network access per environment is None, Trusted (package registries, GitHub, cloud SDKs), Full or Custom, and the default is Trusted. GitHub credentials stay encrypted on Anthropic's servers and never enter the VM; git operations go through a proxy that attaches a scoped credential server side. The docs are candid that even at None the session can still talk to the Anthropic API, which may allow data to leave the VM. Cloud sessions offer only Accept edits, Plan and Auto permission modes, not Manual or Bypass. After a few minutes idle the VM pauses with files saved and can later be reclaimed; reopening rebuilds it with the conversation restored but not background processes.
Moving work between machines is one way from the CLI. claude --cloud with a task description starts a new cloud session that clones your current branch from the GitHub remote, not your local checkout, so unpushed commits are invisible to it. If there is no GitHub remote, or the Claude GitHub App is not installed, Claude Code bundles and uploads the local repository instead (under 100 MB, untracked files excluded, and on macOS, Linux and WSL uncommitted changes to files named like secrets left out). claude --teleport pulls a cloud session, its branch and its history down into your terminal. Auto-fix can watch a pull request and push fixes for CI failures and review comments.
Codex cloud works on the same pattern. You choose Work in, then Cloud, and pick or create an environment; Codex inspects the repositories, installs dependencies and tests the workflow before you publish it. Environments have an install script and a start skill for services, environment variables, network secrets that are substituted into requests to allowed hosts, a Package managers network preset plus extra allowed domains, and Tailscale VPN for private networks. OpenAI's docs list 2 vCPUs, 8 GiB memory and 8 GiB disk on Plus and Edu Plus, and 4 vCPUs, 16 GiB and 32 GiB on Pro, Business and Enterprise. From a terminal, codex cloud exec submits a task to an environment and codex apply brings the resulting diff into your local tree.
Claude Code cloud sessions need a Pro, Max or Team plan, or an Enterprise premium or Chat plus Claude Code seat, a claude.ai sign-in (not an API key, Bedrock or Vertex), and a GitHub account. Organizations with Zero Data Retention cannot use them, and on Team and Enterprise an Owner must enable the GitHub connector and the allow_remote_sessions policy. Codex cloud needs a ChatGPT Plus, Pro, Business, Healthcare, Education or Enterprise plan; Free and Go do not get it, as reported.