GitHub Copilot moves agents into app, cloud, and sandboxes

The center of AI coding is moving from suggestion quality to execution control. On June 2, 2026, GitHub used Microsoft Build to expand the Copilot app technical preview to all existing Copilot Pro, Pro+, Business, and Enterprise customers, while also putting local and cloud sandboxes for Copilot into public preview.

The Copilot app is designed as a desktop home for agent-native software work rather than another editor plugin. Developers can start sessions from issues, pull requests, prompts, or prior sessions, then run parallel agents in separate git worktrees and branches. The app also lets users review plans and diffs, validate behavior in an integrated terminal and browser, and open pull requests that still pass through the team’s normal checks and merge requirements.

The most important interface change is canvases. Instead of leaving progress buried inside a long chat transcript, a canvas gives a work object a visible surface: a plan, pull request, browser session, terminal, release checklist, migration board, incident workflow, spreadsheet, or dashboard. Users can inspect, edit, reorder, approve, or redirect the work there, while the agent reads and updates the same structured state.

That matters because coding agents increasingly act, not just answer. GitHub’s sandbox release gives Copilot isolated places to run commands and touch code. Local sandboxes can be enabled with /sandbox enable, restricting shell commands initiated by Copilot across filesystem, network, and system capabilities. Cloud sandboxes start with copilot --cloud and run as fully isolated, ephemeral Linux environments hosted by GitHub, inheriting existing Copilot cloud agent policies.

The broader Build bundle includes voice input, cloud sessions, cloud automations, Copilot CLI sessions inside the app, agentic browsing for end-to-end UI verification, code review effort tiers, and a generally available Copilot SDK for Node.js/TypeScript, Python, Go, .NET, Rust, and Java. The throughline is clear: GitHub is trying to make Copilot a shared agent runtime across desktop, terminal, cloud, and custom tools.

The next test is adoption inside governed engineering teams. If agents are expected to run builds, browse local apps, edit files, comment on reviews, and merge under conditions, the winning product will need durable audit trails and policy controls as much as strong models. GitHub’s June 2 release is a step toward that operating layer.