What changed
GitHub is pushing its Copilot app beyond the chat box with canvases: small full-stack applications that run inside the desktop app and communicate with the agent in both directions. In GitHub's September 24 explanation, the examples range from a Connect Four game to a local package-management interface, a SQLite workspace, and a workflow surface for coordinating research, prototyping, implementation, and review.
The important detail is that a canvas is not just generated documentation or a richer chat message. It is an interactive surface that can call APIs, run local code, expose controls, and send state back to the agent. GitHub's documentation describes the same model as a shared artifact: people can edit the surface directly while the agent updates it or calls capabilities exposed by it.
Why it matters
Chat is a useful place to state intent, but it is a poor long-term interface for repeated operations. If every change to a board, database, release checklist, or browser session has to become another natural-language turn, the agent becomes both the tool and the UI. A canvas turns that workflow into a reusable product surface.
That changes the economics and the ergonomics of agent work. A person can inspect the current state, use ordinary controls for deterministic actions, and reserve the model for judgment, synthesis, or orchestration. The result is easier to steer and verify than a long transcript, and it can keep human approval in the loop without forcing every interaction through a prompt.
The builder takeaway
GitHub's canvas model gives agent builders a practical test: when a workflow has durable state, structured controls, or repeated human review, build a surface instead of adding more prompt instructions. The app can create a canvas from a natural-language description, then store it either at project scope in .github/extensions for a team or at user scope in ~/.copilot/extensions for personal use. A typical extension contains metadata, an entry file such as extension.mjs, and optional JSON artifacts for persisted state.
This is a product shift worth watching. The next generation of agent applications may be judged less by how well they chat and more by whether they can create the right work surface at the right moment.
Sources
- When chat is the wrong UI — GitHub's September 24, 2026 product explanation and examples.
- Working with canvas extensions in the GitHub Copilot app — official creation, scope, structure, and interaction docs.
- About the GitHub Copilot app — official availability and product context.