GitHub Copilot Canvases
Canonical version: 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
About Sébastien
Ready to get to the next level?
Found this valuable? Share it with someone who needs it.