GitHub launches Copilot desktop app for agentic work

GitHub has introduced a GitHub Copilot desktop application for agent-based software development. The app is in technical preview for existing Copilot Pro, Pro+, Business and Enterprise users.

The desktop app is designed as a central workspace for developers managing multiple AI agent sessions at once. It brings active sessions, issues, pull requests and background automations into a single "My Work" view across connected repositories.

Each session runs in its own git worktree, giving developers an isolated copy of a branch while an agent investigates a bug, works on a backlog item or responds to review feedback. The app manages those worktrees automatically, removing the need for manual branch setup and cleanup.

GitHub has also added Agent Merge, a tool that follows a pull request through checks, reviews and merging. Developers can choose whether Copilot should address failing continuous integration checks, respond to review comments, or complete the merge once set conditions are met.

Control centre

GitHub presented the launch as a response to the growing use of AI agents in software engineering and the management challenges that come with them. Developers are increasingly moving between prompts, planning, coding, review and merging through AI-assisted workflows, but existing tools often leave work spread across multiple windows and interfaces.

The company also pointed to rising activity on its platform. Commits on GitHub have nearly doubled year on year to more than 1.4 billion a month, while GitHub Actions is handling more than 2 billion minutes a week.

That growth is putting more pressure on review and validation. One aim of the new desktop experience is to make agent work easier to inspect, so developers can see what was attempted, what was validated and where human intervention is still needed.

A new feature called canvases sits alongside the app. GitHub described them as shared work surfaces that can display a plan, pull request, browser session, terminal, deployment, dashboard or workflow state, allowing both the developer and the agent to update the same view.

The aim is to move agent output beyond chat threads into a format where actions and decisions are visible in one place. Developers can edit, reorder, approve or redirect work directly within the canvas.

Sandbox controls

GitHub is also expanding how Copilot agents run code by offering both local and cloud sandboxes. The goal is to let agents test and iterate in contained environments without affecting production systems.

In the local version, Copilot runs in an isolated environment on the user's machine with restricted access to files, networks and system functions. Local sandbox policies can be centrally configured and enforced.

In the cloud option, each sandbox runs in an isolated temporary Linux environment hosted by GitHub. Organisations can set their own policies, and users can resume sessions remotely from different devices.

Review changes

GitHub has also updated Copilot code review tools to handle the larger number of pull requests AI agents may generate. Users can now extend code review through custom agent skills, MCP server connections and configurable GitHub Actions workflows.

A medium-tier review option has been introduced, sending pull requests to what GitHub called a higher-reasoning model. Repository administrators can assign either low or medium guidance levels depending on a repository's risk or importance.

GitHub has also added a "/security-review" skill for security-focused checks. A "/rubberduck" skill is now generally available and can use multiple model families to critique an implementation and identify issues.

Support for Copilot code review has also been extended to Azure DevOps, giving users one-click review, inline comments and fix suggestions in that environment.

Broader ecosystem

Outside the desktop app, GitHub said its Copilot SDK is now generally available in Node.js/TypeScript, Python, Go, .NET, Rust and Java. The software development kit gives developers access to the same runtime used by the Copilot app to build internal tools and custom agents.

Copilot CLI has also been updated with a redesigned terminal interface, voice input using on-device speech-to-text, and scheduled tasks. Cloud automations can now let agents run on a schedule, react to GitHub events, open issues and leave comments, with permission requested before each write action by default.

GitHub is also expanding integration with external software vendors through partner-built agent apps embedded in GitHub. Named partners include LaunchDarkly, Bright, Amplitude, Sonar, Endor Labs, Octopus Deploy, Packfiles, PagerDuty and Miro.

One customer, Avanade, said the product gives engineering teams a single place to direct multiple agent-led tasks.

"The GitHub Copilot app is the latest in a line of AI tooling from GitHub that is transforming our business. Moving beyond AI assistance, the app has provided a much-needed control center for agentic development. Our forward deployed engineers can dispatch a cohort of agents and manage multiple initiatives, all from one location. Easy access to plans and autopilots with the ability to run interactive sessions or step into code where needed," said David Jobling, Master Technology Architect, Head of Technology & Delivery Futures, Global Solutioning & Delivery, Avanade.