Three years into the current wave of AI use, chat remains the main interface for large language models. Its flexibility makes it a useful starting point when a product cannot anticipate what users will ask. But once someone knows the task they want to accomplish, a conversation may no longer be the best way to do it.
Academic Steven Pinker has argued that AI’s promise lies in task-oriented applications rather than the first-person chatbot format that initially reached a mass audience. The practical question is how to give users an interface suited to their immediate work.
A canvas as a working interface
In the GitHub Copilot app, a canvas offers one approach: a small full-stack application running inside the app without browser chrome. The Copilot agent can communicate with the application’s server, and the server can communicate back. This combines a conventional software interface with two-way interaction with an agent.
A Connect 4 game illustrates that arrangement. The user plays through the canvas, the canvas communicates with the agent, and the agent controls its side of the game. The demonstration includes GPT-5.6 Sol with high reasoning and a reported win against GPT-5.6 Luna with no reasoning.
Creating this kind of canvas starts with a request describing the game and the interactions it should demonstrate. Because the GitHub Copilot app already understands canvases, users do not need to explain the underlying concept in their prompt.
Build a tool rather than repeatedly asking an agent
Canvases are not limited to displaying web content. As full-stack applications, they can call third-party APIs and execute code locally on the user’s machine.
One example is a Winget interface that searches the package registry and manages local packages, including installation and removal. Its routine operation does not require AI. That distinction is central to the argument: an agent can build a conventional tool once, rather than spend tokens carrying out every subsequent interaction.
The same reasoning applies to routine development actions such as asking GPT-5.6 Sol Max to stage and commit changes. A chat-first interface can encourage users to delegate operations that a dedicated control could handle directly.
A SQLite canvas provides another example. Instead of describing every database operation to an agent, users can work in a purpose-built interface and write SQL themselves, with features such as IntelliSense. For publishing, a canvas could provide a Windows Live Writer-style editing experience for Jekyll posts rather than requiring authors to work in plain Markdown.
Automating a development workflow
Task-specific interfaces can also coordinate larger workflows. The development process presented here follows six stages:
- Research
- Prototype
- Plan
- Implement
- Iterate
- Finalize
In a chat-based workflow, these stages can repeatedly require the developer to return to the keyboard, inspect prototypes, give direction and advance the process. Yet an agent can conduct research and generate prototypes independently before notifying the developer that a review is needed.
A workflow canvas can make those transitions explicit and automate selected steps, allowing the developer to choose how much involvement to retain. The example is presented as a customizable approach, not a universal prescription for working with agents.
The broader argument is that a text box can make useful AI workflows harder to discover. A dedicated interface exposes actions and review points that are less obvious in an open-ended conversation.
The effort required varies. A SQLite canvas may be generated in a single attempt, while the demonstrated workflow canvas took the better part of a day to refine its design and automation. Canvases offer a way to move beyond chat when the task calls for direct controls, reusable tools or structured agent coordination.