By ByteBot
The GitHub Copilot app opened to all paid subscribers on June 2. Coverage has mostly focused on the dashboard redesign and the catchy “agent-native desktop” framing. The actually interesting part is underneath: git worktrees. Every parallel agent session runs in its own isolated branch checkout, which means you can run up to ten agents against the same repository simultaneously without them overwriting each other. GitHub borrowed this mechanism from a git feature most developers have never touched — and turned it into invisible infrastructure.
What Actually Shipped
The Copilot app is a standalone desktop application — not the VS Code extension, not a tab on GitHub.com. It runs natively on macOS, Windows, and Linux. The release timeline: technical preview for waitlisted users on May 14, expansion to all Copilot Pro, Pro+, Business, and Enterprise subscribers on June 2. If you have a paid Copilot subscription, you can download it now.
The app ships with three features worth understanding: the My Work dashboard, Canvases, and Agent Merge. My Work consolidates active agent sessions, open issues, pull requests, and background automations across all connected repositories in a single view. Canvases are bidirectional work surfaces where agents update plans, terminal output, and deployment state in real time — you can redirect them mid-task. Agent Merge watches your CI, handles review feedback, and merges when your conditions are met.
The Git Worktree Architecture
Here is how parallel agent sessions actually work without turning your repository into a merge disaster. Each session runs in its own git worktree — a real, isolated copy of a branch checked out at a separate path on your machine or in a cloud sandbox. The app creates and manages every worktree automatically. No manual setup, no stashing, no branch juggling.
# What the Copilot app handles automatically for each session:
git worktree add ../agent-session-auth feature/auth-agent
git worktree add ../agent-session-perf feature/perf-fix
# Each agent works in total isolationUp to ten sessions can run simultaneously on the same repo. They do not touch each other’s files, conversation state, or task state. Merge conflicts can still emerge when you combine branches — exactly like human developers working in parallel — but that is the correct place to resolve them, not mid-execution. The worktree approach stops the common failure mode of parallel AI agents overwriting each other, which has been a persistent problem in agentic coding setups.
The Workflow That Changes
The shift the Copilot app embodies is from pair programming to task assignment. In the pair programming model, you stay at the keyboard and the AI suggests what to type. In the assignment model, you open an issue, delegate it to an agent, and review the resulting pull request. The app is built for the second model. You drop into My Work, pick a few issues, spin up sessions, and come back to PRs.
This is also GitHub’s strategic response to Claude Code and OpenAI Codex gaining ground with terminal-native agentic developers. By building the orchestration layer around GitHub’s own primitives — issues, PRs, Actions, code review — GitHub is betting you will stay in its ecosystem rather than switch to a standalone terminal agent. The integration advantage is real for teams already running everything through GitHub.
The Billing Problem
There is a tension that deserves more attention. GitHub switched to token-based billing on June 1 — one week before the app’s full rollout. One AI credit equals $0.01. Copilot Pro subscribers get $10 per month, which is 1,000 credits. A single agentic session routinely runs $30 to $40 in credits. Running ten parallel sessions means ten parallel credit burns. Developers running heavy agentic workloads are already reporting 10× to 50× cost increases compared to the previous billing model.
The Copilot app is architecturally designed to encourage running more agents simultaneously. The billing model punishes exactly that behavior. Pro-tier users who treat the $10/month budget as a hard cap will need to be selective. Pro+ subscribers at $39/month get 3,900 credits — enough for serious work, but not unlimited parallel sessions. Plan accordingly before you start queuing up ten jobs at once.
The Multi-Agent Angle
The Copilot app is also the desktop home for GitHub’s Agent HQ, which lets you run Claude, Codex, and Copilot on the same repository simultaneously. Different agents handle different workstreams, and worktrees keep their outputs isolated. For teams that live inside GitHub’s pull request and issue workflow, this is the strongest case for the app: multi-vendor AI orchestration without leaving the tools you already use.
According to Anthropic’s 2026 Agentic Coding Trends Report, 80% of developers now use AI coding agents in their workflows. Trust in AI accuracy has dropped from 40% to 29% year over year. That tension — mass adoption paired with declining confidence — is exactly why features like worktree isolation, Agent Merge review gates, and Canvas visibility exist. The tooling is trying to keep humans in the loop while agents do more of the work. Whether GitHub’s approach is better than the alternatives depends on where your workflow already lives.
