# GitHub Copilot Canvases Canvases are durable, shared visual surfaces in the [[GitHub Copilot App]] where you and an AI agent operate on the same plan. Instead of workflow state living buried in chat history, a canvas makes it explicit and persistent: phases, decision points, validation gates, and drafts update live as the agent works. Shipped at the app's GA (June 2026); created with the `/create-canvas` command or installed as extensions from [[Awesome GitHub Copilot]]. GitHub's pitch is that canvases make agentic workflows: - **Visible**: operational state is inspected directly on the canvas, not reconstructed by parsing a conversation - **Steerable**: humans guide execution between checkpoints, approve transitions, and redirect work without losing context - **Cost-efficient**: building a canvas costs credits up front (2,000-3,000 per canvas in GitHub's own experience), but it reduces repeated prompting, context loss, back-and-forth, and rework over time Architecturally, a canvas is a thin steering layer: it dispatches prompts to the Copilot agent and renders the results ("the agent does the work; the canvas reflects the result"). Reference examples: Site Studio (section-by-section website authoring) and Java Modernization Studio (assessment → remediation → validation → ship), both on [[Awesome GitHub Copilot]]. This is the same direction as artifact-centric agent UIs elsewhere: move the human-agent contract out of the chat transcript and into a structured, persistent surface. ## References - How canvases make agentic workflows visible, steerable, and cost-efficient (Ayan Gupta): https://github.blog/ai-and-ml/github-copilot/how-canvases-make-agentic-workflows-visible-steerable-and-cost-efficient/ - Canvas extensions catalog: https://awesome-copilot.github.com/extensions/ ## Related - [[GitHub Copilot App]] - [[GitHub Copilot]] - [[Awesome GitHub Copilot]] - [[AI Agents]] - [[AI Agent Harness]]