GitHub has introduced a new Copilot app designed to serve as a desktop control center for software development workflows built around AI agents. Announced at Microsoft Build, the app is meant to help developers manage multiple agent tasks at once, from debugging and implementation to code review and merging.
The company says the new experience reflects a broader shift in how software is being built as agents take on more of the day-to-day work. Instead of treating AI as a simple chat assistant, GitHub is positioning Copilot as an environment where developers can direct, monitor and verify autonomous work across repositories.
The Copilot app is now available in technical preview for users on Copilot Pro, Pro+, Business and Enterprise plans.
GitHub says the app centers on a "My Work" view that surfaces active sessions, issues, pull requests and background automation in one place. The idea is to reduce the context switching that often comes with managing several agent tasks at the same time. Each session runs in its own git worktree, giving every agent an isolated copy of the branch so parallel work does not interfere with other tasks.
That setup is intended to remove much of the manual branch handling and cleanup developers typically face when juggling multiple lines of work. The app also connects to existing issues, pull requests and repositories to give agents the context they need to begin work.
GitHub is also adding features aimed at making AI work easier to inspect and steer. One of those is canvases, which the company describes as bidirectional work surfaces for people and agents. A canvas can show plans, pull requests, browser sessions, terminals, deployment status or workflow state. As agents update the canvas, developers can edit, approve, reorder or redirect that work on the same surface.
The company frames canvases as part of what it calls agent experience, or AX, which is intended to give humans a clearer way to collaborate with automated systems than a long chat thread can provide.
To further limit risk, GitHub is pairing the app with cloud and local sandboxes. These environments let agents run code and test changes without touching production. Local sandboxes run on a developer’s machine with restricted access, while cloud sandboxes run in isolated Linux environments hosted by GitHub. Organizations can set policies for either option, and cloud sessions can be resumed from different devices.
GitHub is also extending automation around pull requests. Its Agent Merge feature can monitor CI, watch for reviewer approvals, respond to failing checks and continue the process until conditions set by the user are met. Developers decide how far the automation can go, including whether it should fix failing tests, address feedback or merge code.
The company says its review tools are evolving as agent output increases. Copilot code review now includes a medium review tier that routes work to a more capable reasoning model for better precision and recall. Administrators can assign review levels by repository. GitHub is also adding a security-focused skill and a general-purpose critique skill called /rubberduck, which is now generally available.
Beyond the app itself, GitHub is expanding the underlying platform. The Copilot SDK is now generally available across several languages, including Node.js, Python, Go, .NET, Rust and Java, so teams can build custom tools on the same runtime. Copilot CLI has also been updated with a new interface, voice input and scheduled tasks.
GitHub says the changes are part of a broader effort to make its platform the home for developers and their agents as agentic workflows become more common across repositories, pull requests and API usage.