GitHub is expanding the Copilot experience with Canvases, a new feature designed to make complex AI-agent workflows easier to monitor, steer and approve. Announced on August 17, 2026, the feature moves long-running agent tasks away from chat-only interfaces and into persistent workspaces where workflow state remains visible.
According to GitHub Developer Advocate Ayan Gupta, chat remains useful for defining intent and refining ideas, but becomes harder to manage once execution generates logs, decisions, corrections and review points. Canvases are intended to reduce that complexity by keeping progress and human intervention points visible.
Instead of repeatedly prompting an agent and reconstructing previous context, developers can organize work into defined stages.
Canvases can maintain persistent workflow states, highlight decision points, show live progress and preserve drafts as agents continue working. They can also introduce approval gates so developers remain involved before important actions proceed.
Gupta explained that the design turns Copilot from a “prompt-by-prompt conversation into a workflow with a memory, structure, and agency.”
Developers can also use the ‘/create-canvas’ command to build reusable workflows around recurring development tasks.
GitHub highlighted Java Modernization Studio as one example of the approach. Application modernization often involves assessment, remediation, validation and deployment. Instead of keeping all these steps buried inside a conversation, the canvas can display the current modernization phase, blockers, assessment results and tasks awaiting approval.
According to Gupta, this allows human reviewers to spend more time on important technical judgments while agents continue execution between checkpoints.
However, creating sophisticated workflows can have an upfront AI cost. Gupta reported using roughly 3,000 AI credits to create the Java modernization canvas.
Another example, Site Studio, applies the same model to creating and managing website content.
Individual sections can move through drafting, editing and review stages while retaining their current status. This prevents revisions and feedback from becoming scattered across long chat histories.
Gupta said approximately 2,000 AI credits were used to build Site Studio.
GitHub recommends starting with a small recurring workflow, defining its stages and approval points, and refining the canvas through actual use.
The company argues that while creating a canvas may initially consume significant AI credits, repeated use could reduce context rebuilding, redundant prompting and rework.
Also Read: Gemini 3.7 Flash Arrives: Why Google Is Betting on Faster AI for Agentic Workflows
As AI agents take responsibility for increasingly complex development tasks, simply generating code is no longer enough. Developers also need visibility into what agents are doing, what has been validated and where human approval is required.
Canvases represent GitHub's attempt to make AI-assisted development more persistent, auditable and controllable turning Copilot from a conversational assistant into a structured collaborator for multi-stage workflows.