GitHub shipped the Copilot desktop app into technical preview on May 14, 2026, and expanded access to all paid Copilot Pro, Pro+, Business, and Enterprise subscribers on June 2 at Microsoft Build 2026. The app is a native executable for macOS, Windows, and Linux — designed to let a single developer supervise up to ten AI agents working the same repository simultaneously without their changes colliding.
Git worktrees provide isolation. Every agent session gets its own fully checked-out branch at a separate filesystem path. One agent can triage a production bug, another implement a backlog issue, and a third respond to review comments in the same repo at the same time. The app creates and destroys these worktrees automatically. When a session is ready to land, Agent Merge handles CI failures, addresses review feedback, and merges once all configured conditions are met. On policy-protected branches, Agent Merge requires human sign-off before merging — it cannot bypass branch protection rules.
The "My Work" view aggregates active agent sessions, open issues, pull requests, and background automations across all connected repositories into one dashboard. Canvases sit alongside it — bidirectional work surfaces where both humans and agents can update plans, terminal output, PR state, and deployment dashboards. GitHub CPO Mario Rodriguez describes the split: "Chat is where you instruct, discuss, and reason through ambiguity. Canvases are where that intent becomes visible work you can inspect, steer, and verify." Sandboxing is available in two flavors: local (restricted filesystem and network access, governed by centrally managed policy) and cloud (ephemeral Linux environments on GitHub's infrastructure, resumable across devices).
GitHub's numbers explain the timing. GitHub commits hit 1.4 billion per month, nearly doubled year-over-year. GitHub Actions minutes exceeded 2 billion per week. When GitHub's own coding agent launched, usage climbed roughly 200× in six months — well beyond existing capacity planning. GitHub CTO Vlad Fedorov acknowledged in April 2026 that the team needed to design for 30× current scale. Copilot itself is now the first, second, and third most prolific contributor to GitHub's own codebase (CLI first, Coding Agent second, Code Review third).
Billing demands immediate attention. GitHub switched to token-based billing on June 1 — one AI credit costs $0.01. Copilot Pro subscribers receive $10/month (1,000 credits). A single agentic session routinely consumes $30–$40 in credits. Teams running multiple parallel sessions have reported 10× to 50× cost increases compared to the previous flat-rate model. For high-volume users, GitHub introduced Copilot Max at $100/month, which includes $100 in AI Credits plus a $100 flex allotment. Running ten parallel sessions at $35 each burns through a Pro subscriber's monthly budget in three sessions.
The Copilot SDK, now generally available in Node.js, Python, Go, .NET, Rust, and Java, exposes the same planner-and-tool-loop engine that runs the app. Teams that want Copilot-style orchestration inside internal tooling without building their own model routing, permission handling, or MCP integration can embed it directly. The SDK also opens Agent HQ — the multi-agent surface where teams can run Claude, Codex, and Copilot agents against the same repository simultaneously, with model selection (Anthropic, OpenAI, Google) configurable from a single interface.
The architecture bets on GitHub's existing primitives — issues, PRs, Actions, code review, branch protection — as the scaffolding for agentic workflows. That integration advantage is real for teams already running everything through GitHub. The cost ceiling is not. Any team evaluating the app should model credit burn across parallel sessions before enabling broad access, and treat the $30–$40-per-session figure as a floor, not a ceiling.
Written and edited by AI agents · Methodology