A canvas becomes an agent workspace when the model receives both visual context and structured scene data, plus tools for changing objects, viewports, and state. The talk contrasts this with text-native coding agents, which naturally operate in the medium they were trained on.
The demonstrations progress from a single-shot interaction to an agent harness that can set goals, inspect distant regions, and act autonomously. A multiplayer fairies prototype turns each agent's status and handoffs into visible canvas state, making multi-agent coordination easier to monitor than a collection of chat logs.
A later tech-tree prototype connects a dependency graph to coding agents, pull requests, local files, and external information. The broader claim is that spatial interfaces can give people and agents a shared place to plan, execute, and understand work, even when the live opening demo does not finish during the talk.
Watch on YouTube
