GitHub introduced canvases in the Copilot app on September 25, 2026, presenting a workflow-first way to design an interface. A canvas extension is a customizable surface shared by a person and an agent. It can be a kanban board, issue-triage board, release checklist, dashboard, form, or spreadsheet.
Unlike a chat that only displays an answer, a canvas is bidirectional. The agent can update cards and data as work progresses, while the user can change the same surface through buttons, filters, cards, and other controls. GitHub describes it as a live shared whiteboard: the interface becomes part of the work rather than merely its output.
Creating a canvas does not require hand-written UI or layout work. In an agent session, the user runs `/create-canvas` and describes the workflow the canvas should support, the controls the user needs, and the actions the agent may take. The agent opens a first version in a side panel, which can then be refined through conversation and use.
This model fits work that benefits from visible state, such as tracking release notes, triaging issues, managing a launch checklist, or exploring data. The value is not only faster interface creation; it is also that the agent’s progress becomes something a person can inspect, operate, and correct directly.
A generated interactive surface is not automatically a reliable workflow. Teams still need to review its data sources, update permissions, side effects, versioning, and audience. If a canvas connects to issues, repositories, or deployment systems, approvals, tests, and change records remain necessary even when the interface feels simple.
GitHub also points to community canvas extensions for release notes, kanban, and issue triage. Reusing one can shorten the starting point, but teams should inspect third-party permissions, instructions, and data flows. An extension that fits one workflow may not fit another repository or sensitive dataset.
The direction is a move from an AI agent in a chat box to an agent in a persistent workspace. A safer adoption path is to start with a small checklist or read-only dashboard, verify data and action boundaries, and only then allow agent updates or writes to external systems.
