When chat is the wrong UI
GitHub introduces canvases in its Copilot app, replacing chat interfaces with customizable, full-stack tools for AI-driven tasks like games or package management.
GitHub argues that while chat interfaces with AI models have been the default, they are often inefficient for specific tasks. The company proposes canvases—customizable, full-stack mini-applications within the GitHub Copilot app—as a more practical alternative. These canvases operate without browser chrome and enable bidirectional communication with AI agents, allowing for more direct and task-oriented interactions. The shift aims to reduce unnecessary token usage and improve workflow efficiency by letting users build tools tailored to their immediate needs.
A canvas can function as a standalone application, such as a Connect 4 game where users play directly against the AI within the Copilot app. This example demonstrates how canvases can handle interactive tasks without relying on chat-based instructions. The GitHub Copilot app recognizes canvases as full-stack applications, enabling them to call third-party APIs or execute local code. This flexibility allows users to perform actions like managing software packages via Winget without repeatedly querying the AI agent.
The introduction of canvases addresses the inefficiency of using chat interfaces for repetitive or structured tasks. For instance, instead of repeatedly asking an AI to stage and commit code changes, users can create a canvas tool that performs these actions locally. This approach conserves tokens and reduces latency, as the tool handles routine operations independently of the AI agent. The concept aligns with the idea that AI should augment workflows rather than serve as the sole interface for every task.
GitHub highlights that canvases empower users to take direct control over their interactions with AI. By building custom UIs for specific tasks—such as querying an SQLite database—they can leverage intellisense and other features without relying on chat. This shift reflects a broader trend toward more practical and user-driven AI integration, where the interface adapts to the task rather than forcing the task into a chat-based model.