The GitHub Copilot desktop app went from technical preview to a real product at Build 2026, and the pitch is now explicit. This is not an IDE. It is the orchestrator for the seven agents you are pretending to manage while you are actually in a meeting. Windows, macOS, and Linux clients, three operating modes (Interactive, Plan, Autopilot), and a mobile handoff that lets you check in on agent runs from GitHub Mobile while you are walking the dog.
The interesting engineering bit is that every Copilot agent session now runs inside its own git worktree, auto-provisioned, auto-cleaned, no manual branch juggling. If you have ever tried to run three Claude Code or Codex sessions against the same repo in parallel and watched them stomp on each other’s working directory, you understand why this matters. Worktrees are how git was always supposed to handle parallel work, and the dev tooling has just never made it ergonomic enough for humans to bother with. Agents, it turns out, do not mind the ergonomics. They just need the isolation.
Agent Merge is the headline feature for procurement decks. After an agent opens a PR, the Merge mode keeps driving: it watches CI, addresses failing checks, responds to reviewer feedback within configured guardrails, and merges when the conditions you defined are satisfied. You can dial the autonomy down to “just keep CI green and ping me” or up to “merge it when the lights turn green.” This is the first piece of mainstream developer tooling that ships with explicit, configurable autonomy levels for the merge step, which has historically been the one place humans insisted on staying in the loop.
The strategic frame is that GitHub is no longer competing on “best autocomplete” against Cursor and Codex. It is competing on workflow surface area. The desktop app is the layer where a developer who has eight agents working on eight tickets can actually see what is happening. That is a different product than an editor extension, and it is harder for Cursor and Windsurf to replicate from inside an IDE that was designed for one human at a keyboard. Whether the workflow lock-in is enough to keep developers paying for Copilot in a market full of cheaper agents is the question the next four quarters will answer.