Inside GitHub Copilot's New Desktop App: Agentic Development Goes Native

Parallel sessions with three control modes

The app lets you run multiple agent sessions simultaneously, each with pause and resume support. Developers choose how much autonomy to grant each session through three distinct modes :

• Interactive: The agent offers step-by-step guidance and waits for approval at each stage.
• Plan: The agent drafts a plan of changes, which the developer must review and approve before any code is written.
• Autopilot: The agent works fully autonomously on its assigned branch, writing, testing, and iterating without manual checkpoints.

This sliding scale of control is designed to accommodate a range of confidence levels. A developer might set a well-understood bug fix to Autopilot while staying in Interactive mode for a sensitive architectural refactor.

Agent Merge: the PR lifecycle on autopilot

Perhaps the most debated feature of the release is Agent Merge, an automated pull-request lifecycle manager. Once a PR is created, Agent Merge can respond to reviewer comments, diagnose and fix failing CI checks, and automatically merge the PR once all conditions are met—closing the loop from issue to merged code without human intervention .

This does not mean merge happens secretly. Conditions must be explicitly satisfied, such as passing CI and receiving required approvals. The agent can respond to reviewer feedback and push fixes, but the guardrails of branch protection rules and code review still apply .

Canvases: a visual workspace for steering agents

Moving beyond a purely chat-based interaction model, the desktop app introduces Canvases—visual workspaces where plans, pull requests, terminal outputs, and deployment statuses appear as interactive surfaces visible to both the developer and the agent simultaneously .

A Canvas gives a shared visual grounding point. For example, a kanban board Canvas might display all cards while the agent moves items and the developer drags to reprioritize. A review Canvas might expose a list of open questions, acceptance statuses, and comment threads. This is a deliberate move to reduce the cognitive overhead of tracking agent work through long chat threads alone .

Voice dictation: hands-free agent management

The desktop app includes built-in voice dictation powered by on-device speech-to-text . After configuring a keyboard shortcut and downloading a local transcription model, a developer can dictate prompts to any agent session. No audio leaves the machine, addressing privacy concerns for teams in sensitive environments .

Voice dictation launched simultaneously across the Copilot CLI, which also received rubber-duck mode and prompt scheduling in a major refresh on June 2 . The combined voice capabilities across both surfaces are positioned as a critical enabler for agentic workflows; when managing multiple parallel agents, speaking a prompt can be faster and less disruptive than switching contexts to type.

The billing connection: how the AI Credits model fits

The timing of the desktop app expansion is no coincidence. On June 1, 2026, GitHub retired its premium-request-unit billing system and replaced it with GitHub AI Credits, a metered model where the cost of a Copilot interaction depends on the model used and the number of tokens consumed, including input, output, and cached tokens .

Subscription price points are unchanged: Pro stays $10/month, Pro+ $39, Business $19/user, and Enterprise $39/user. But each plan now comes with a monthly allowance of credits at a rate of 1 credit = $0.01 USD, and when credits are exhausted, the service stops rather than degrading . Code completions and Next Edit Suggestions remain unmetered, but chat, Agent Mode, code review, and tool calls all bill against the credit pool .

The new desktop app, with its heavy autonomous workloads, directly drives token consumption and therefore credit usage. The product expansion gives subscribers the primary interface for consuming those credits, and the billing model is now structurally aligned with the cost of running many long-lived, context-heavy agent sessions .

The strategic arc: from copilot to orchestration platform

The desktop app completes a multi-year trajectory for GitHub Copilot. The original promise was an AI pair programmer inside the editor. The new reality is an execution platform for an entire software development lifecycle—from issue triage through code generation, CI verification, review, and merge—with the developer acting as orchestrator rather than sole operator.

Voice input removes friction in interacting with parallel agents. Agent Merge handles the mechanical follow-through on PRs. Canvases provide a shared visual surface for humans and agents. The My Work dashboard surfaces everything in motion. These are not standalone features; they are the interface layers of a system designed around the assumption that agents are the new primitives of development work, not just chat partners .