GitHub ships the Copilot desktop app at Build 2026, a control center for parallel agent work
GitHub announced a standalone Copilot desktop app at Microsoft Build 2026 on 2026-06-02. It runs every agent session in its own git worktree, ships an Agent Merge primitive for autonomous PR shepherding, and bets on a single My Work view for multi-agent orchestration.
GitHub announced a standalone Copilot desktop app at Microsoft Build 2026 on 2026-06-02, available in technical preview on Windows 11, Windows 11 on Arm, Mac, and Linux for Copilot Pro, Pro+, Business, and Enterprise subscribers. It is the first time GitHub has shipped Copilot as something other than an IDE plugin or a browser surface, and the design choices look pointed at the same problem Claude Code and Cursor have been carving at: how do you direct three or four agents at once without losing track of which one is doing what.
What shipped
The Copilot app collapses active agent sessions, GitHub issues, pull requests, and background automations into a single My Work view. Each session runs inside its own isolated git worktree, which means multiple agents can be working on the same repository at the same time without branch collisions, manual setup, or cleanup after the fact. GitHub describes worktree management as something the app handles end to end, including teardown.
Agent Merge is the standout primitive. Once a pull request is in flight, Agent Merge will monitor CI checks, track required reviewers, respond to review feedback, and complete the merge under conditions the developer specifies. Autonomy is a slider, not a default: drive CI back to green, address review comments, or merge once everything clears. It is the first vendor surface that treats the post-draft phase of a PR as a discrete agent target rather than a notification center.
Canvases are the agent-experience side of the launch, structured interactive surfaces that sit over a work object so an agent and a developer can edit, inspect, and steer the same artifact in the same view rather than in chat-only back-and-forth. The Copilot SDK reached general availability the same day in Node.js, TypeScript, Python, Go, .NET, Rust, and Java, which gives the agent and Canvas surfaces a real path for third-party tooling.
Around the headline pieces, the app picked up on-device voice control (speech- to-text runs locally, no audio leaves the machine), cloud-hosted agent sessions, scheduled cloud automations that run while the machine is asleep, and agentic browsing in an integrated browser so the agent can click and screenshot to verify its own UI changes. Cross-surface continuity lands at the same time: a session started in the desktop app can be resumed from GitHub Mobile or github.com via remote session control, now generally available.
The code-review surface also moved. Copilot review now offers a medium tier that routes pull requests through a higher-reasoning model, with per-repository admin controls to assign lighter models to low-risk paths and reserve the medium tier for high-impact code. /security-review is a new code-review skill aimed specifically at security; /rubberduck is now generally available for cross-model critique.
Where this lands in the market
The shape of the launch is GitHub catching up to a pattern that Anthropic’s Claude Code, Cursor, and Cline have all been pushing for the last six months: an agent-first standalone surface that is not welded to a specific editor. GitHub’s wedge here is the data layer. The My Work view does not have to assemble its inputs from scratch. Pull requests, issues, repositories, and CI runs are already first-class GitHub objects, and Agent Merge plugs into a code-review and checks-status flow GitHub already owns. The standalone editors get to build a beautiful agent surface and then graft a code-host integration on top; GitHub gets to start from the code-host side and build the agent surface in front of it.
That is the durable read. The timing read is sharper. GitHub moved Copilot to usage-based AI Credit billing on 2026-06-01, and the community reaction is still acute. Shipping the desktop app and a paid-tier-visible Agent Merge surface twenty-four hours later is not coincidence. The implicit pitch to the Pro and Pro+ users who just watched their projected monthly bill jump is that the new spend buys real agent orchestration capacity, not just more inference. Whether that pitch holds depends entirely on whether Agent Merge and parallel worktree sessions land smoothly in the preview cohort or surface the same kind of “credits drained in two days” complaints currently filling the GitHub discussion thread.
The other competitive read is on the editors. VS Code remains the universal Copilot host, and the desktop app does not retire it. But a substantial chunk of what made Cursor attractive (parallel agent sessions, a workspace-level command surface, a single place to see what every agent is doing) is now in the official GitHub product. The question is whether the IDE-switching tax for Cursor users falls faster than the Copilot-app’s preview-grade rough edges get smoothed out.
What’s worth watching
-
Whether Agent Merge defaults match developer expectations. The slider design is deliberate, but defaults set behavior. If the default tier of Agent Merge starts auto-completing merges without visible review intervention, expect the same backlash that hit automatic dependency updaters a few cycles ago. If it ships closer to “drive CI green, then stop”, trust accrues slowly but cleanly.
-
Whether parallel worktree sessions create new billing surprises. A parallel-agents UX paired with token-based pricing creates a new cost shape: a developer running four agents at once draws from the credit pool four times as fast. The first published blow-up budget post will set the cohort-level read on whether the desktop app is a productivity step up or a faster path to the monthly cap.
-
The SDK’s adoption curve. Copilot SDK GA across seven languages is a real surface, but third-party tooling built on top has historically lagged GitHub Marketplace numbers. The first genuinely novel community-built Canvas or Agent Merge skill will be the data point worth tracking.
The follow-up beat is around 2026-09-01, after the technical preview opens up further and the first wave of paid-tier usage data lands.