VS Code lets Copilot agents keep working after you close the laptop

GitHub has updated Copilot in Visual Studio Code with cloud agent sessions that keep running after the laptop is closed. In its Universe changelog post, the company says one session can be resumed from the editor, a phone, the web or the terminal.

One session that follows you across clients

A cloud agent session starts from the Agents window in VS Code and runs on GitHub’s infrastructure. According to GitHub, you can reopen it in the editor, in the Copilot desktop app, on your phone through GitHub Mobile, on the web at github.com, or from the terminal with Copilot CLI. Sessions begun in a Slack or Microsoft Teams direct message can be picked up in those same places.

For people who prefer to keep code on their own hardware, a new Agent Host separates Copilot sessions from individual VS Code windows, so work continues after a project folder or window is closed. Agents can also run on a separate machine through Remote SSH and dev tunnels. GitHub notes that the remote machine and its Agent Host must stay running.

The phone can act as a remote control for a development machine too. From a project folder, a developer runs code agent host --github-environment in a terminal, then checks progress or starts tasks from GitHub Mobile. The agent still does the work on the computer, not on the phone.

GitHub also lists several other changes:

• The Copilot harness, built on the Microsoft Copilot SDK, becomes the default for new sessions in the editor and the Agents window.
• Copilot CLI can be installed with the VS Code installer on Windows, when adding VS Code to the path on macOS, or by running copilot on Linux.
• A preview lets users run several models and harnesses in parallel, compare results side by side, ask a Judge model for a recommendation, or choose Synthesize to merge ideas from the attempts.
• The open Agent Host Protocol (AHP) lets third parties build clients on the same live sessions, with libraries for TypeScript, Rust, Go, Kotlin and Swift.

Computer use and connectors arrive as previews

Computer use is experimental. On macOS or Windows it is switched on with the setting chat.agentHost.computerUse.enabled, followed by a restart and the command /computer on in a session. GitHub says each application needs the user’s approval and that organizations can disable the feature by policy. On macOS the setup is shared with Copilot CLI and the Copilot app, so system permissions are granted once.

The Customization Marketplace, which gathers skills, MCP servers and plugins under a Browse page, is in preview and on by default. Copilot connectors for services such as Slack, Notion and Microsoft Learn are also in preview, on all Copilot plans, and work with the Copilot harness. Business and Enterprise administrators must allow Editor preview features before either appears. Administrators can also require or block named plugins and restrict installs to approved marketplaces; a repository file, .github/copilot/settings.json, can register marketplaces, though GitHub stresses these are defaults and not enforcement.

What changes for developers and admins

The practical shift is that an agent run no longer depends on one window staying open. Until now, a long task meant leaving an editor session alive; with the Agent Host, closing a folder does not end the work, and a cloud session survives a closed lid. The trade-off is that remote and phone-controlled setups need a machine left running, which turns a developer workstation or a dev box into infrastructure someone has to keep up.

The release lands alongside other Universe news. The same day, VS Code shipped version 1.144.0, and GitHub published a sibling Copilot app update. Our earlier coverage of the Copilot app’s computer use mode shows the feature now exists in both the editor and the app, so admins have to govern it in both places.

Cost is the other open question. We reported that GitHub Enterprise now includes Copilot for every licensed user, with AI billed by use, and that Copilot Fusion offers a 30% discount for now. Long-running background agents and parallel model comparisons are exactly the kind of work that consumes usage, but the changelog does not say how cloud sessions, the Judge model or parallel runs are metered.

What the changelog leaves unsaid

The post gives no performance figures, no adoption numbers and no list of which plans include cloud sessions, so the claims about consistency and convenience are GitHub’s own and unverified. The Judge and Synthesize comparison, the marketplace and connectors are previews, and computer use is explicitly experimental. Details on AHP and the Agent Host are deferred to a separate deep dive and to the VS Code release notes. Developers who want to try the phone workflow need to run the host command themselves; nothing is enabled automatically.