Software development is evolving from static code to orchestrating live systems where developers and AI co-create in real time. The GitHub Copilot App Canvas, demonstrated by the Agent Runtime canvas extension, facilitates this shift in software building. Canvas is not for end-users but for developers and AI to observe and shape software while it is running. It serves as a shared surface for humans and AI agents to interact with a live model of the system.The Agent Runtime canvas provides observability and control for multi-agent systems during their design and evolution. It visualizes a single SystemModel that both humans and AI agents can edit simultaneously. This canvas displays requirements, active agents, task flows, intermediate artifacts, validation results, live system state, and a change timeline. These panels are intended for development observability, not for end-user production interfaces.AI agents interact with the system by calling five specific actions: decompose, execute, validate, update design, and track state. Human controls and these AI actions funnel through the same shared store, creating a unified interaction model. The core of the demo is the SystemStore, which broadcasts every mutation to all connected panels, ensuring a single source of truth.Human controls and AI actions map to the same underlying store methods, meaning neither the human nor the AI is privileged. The system's execution advances task graphs visibly, showing the lifecycle of each task and its active agent. Failure propagation is also observable, with downstream tasks being marked as blocked if a dependency fails.Validation is a first-class feature, with tests asserting properties of the live system model rather than static code. This creates a continuous, visual feedback loop of decompose, execute, validate, redesign, and re-validate. Running the demo requires a GitHub Copilot CLI with canvas support and cloning the provided repository. The extension auto-discovers from the local workspace and can be initiated by asking Copilot to open the Agent Runtime canvas with a specific requirement.The primary use case for Canvas is observability in agentic systems, which are notoriously difficult to debug. The Canvas surfaces intermediate states, coordination order, and failure cascades that production UIs typically hide. This reframes Canvas as a crucial tool for understanding and steering AI actions as teams adopt agent tooling. The architecture of the demo, with its shared, versioned store, supports multi-user collaboration, presenting an open research question for future development.