From scattered chats to an AI agent desktop
GitHub’s Copilot app is a dedicated desktop AI coding environment that centralizes agents, code, and collaboration into a single interface so developers can coordinate multiple assistants across repositories without juggling scattered chat windows. Instead of living inside browser tabs or editor sidebars, the new GitHub Copilot app runs as an AI agent desktop, functioning like an operating system layer that manages parallel agent sessions, background automations, and code review workflows. Announced at Microsoft Build 2026, it introduces a unified “My Work” view that brings together ongoing agent runs, issues, pull requests, and long-running tasks into one dashboard. Each agent session is isolated in its own Git worktree, allowing several agents to work on the same project without stepping on each other’s changes. GitHub is offering the app in technical preview on major desktop platforms for Copilot subscribers, with a waitlist for free tier users.
A unified OS for multi-agent development workflows
The Copilot app reframes GitHub as an AI agent operating system, where developers manage fleets of agents instead of a single coding assistant. The “My Work” dashboard turns what used to be fragmented chat threads into a project-centric view of active automations, reviews, and experiments. Because every agent session runs within an isolated Git worktree, teams can spin up parallel agents on the same repository to explore different fixes, refactors, or feature branches at once. This structure directly supports multi-agent development patterns, where one agent might triage issues while another handles refactoring and a third focuses on documentation. According to Paul Thurrott, the app now works as a “central dashboard for managing AI agents and interacting with GitHub,” marking a shift from occasional AI help toward an always-on layer that coordinates many autonomous workers around a codebase.
Canvases: where agent output becomes shared work surfaces
To move beyond plain chat, the Copilot app introduces canvases—bidirectional work surfaces where agents and humans collaborate on visible artifacts instead of messages alone. A canvas can show a feature plan, a pull request, a browser session, terminal output, or even deployment state. As agents work, they update the canvas directly; developers can then edit, reorder, approve, or redirect steps on the same surface, turning AI suggestions into a structured workflow. GitHub describes this as the beginning of “agent experience (AX)” in the Copilot app, where chat is used to reason about tasks and canvases show concrete progress. This design helps teams treat AI agents as first-class contributors whose work is inspectable and auditable, reducing the risk of hidden changes and making multi-agent development less opaque for reviewers and leads.
Local and cloud sandboxes as the new AI coding environment
Under the hood, the Copilot app reshapes the AI coding environment by standardizing how agents run code. Local sandboxes execute on the developer’s machine with limited filesystem and network access, controlled through centrally enforced policies so organizations can set boundaries on what agents can read or modify. For flexible, device-agnostic workflows, cloud sandboxes run in isolated, ephemeral Linux environments hosted by GitHub, allowing agents to continue work across devices without sharing local state. This separation gives teams more confidence assigning higher autonomy to agents, since experiments and fixes are contained until reviewed. It also prepares workflows where human developers orchestrate a mix of local and cloud tasks—heavy CI-like tasks in the cloud, quick iterations locally—while the Copilot app tracks everything as part of a unified agent-native desktop experience.
Reshaping team workflows, reviews, and third‑party automation
The Copilot app also rethinks how teams structure AI-assisted workflows around pull requests and reviews. Agent Merge can carry pull requests through review, CI checks, and merging, watching for passing checks, required reviewers, and failures. Teams choose how much autonomy to grant, from driving CI back to green to merging when conditions are met. Copilot’s new “medium” review tier routes critical pull requests to a higher-reasoning model, while repository guidelines can keep low-risk code on a lighter setting and send important services through deeper analysis, including a /security-review path. The /rubberduck skill now critiques implementations across models, and Azure DevOps users get native Copilot review with inline comments and fix suggestions. With the Copilot SDK generally available across major languages and partner-built agent apps from LaunchDarkly, Amplitude, Sonar, PagerDuty, and Miro, the desktop shifts into a programmable AI agent hub tied directly into the broader DevOps toolchain.
