A canvas shows you state and still cannot tell you the work is right

Max Drake of tldraw gave a talk published September 10 about putting agents on an infinite canvas, and he is honest about the hard part. Models write code well because text in and text out is the medium they were trained in. They are bad at 2D space. Teaching one to read a canvas from a screenshot plus the underlying data, and to predict what its own actions would do to the layout, was engineering work rather than a clever prompt.

What sits on top is more interesting than the drawings. His agents set their own goals and move their own viewport to go find things, the way a coding agent searches a repo for a definition. They are multiplayer, they have hats and colors, and he defends that as function rather than decoration: with ten agents running, you look at the picture instead of reading ten chat logs. His internal tech tree is a dependency graph where every node is a coding agent you can start, which is the same board tldraw draws by hand when a launch gets close and everyone abandons the task tracker.

The vault’s harness page lists what a loop needs: a goal, a bounded action lane, a state surface the agent can inspect, an evaluator, a repair path, and a stop condition. A canvas is an unusually good state surface. It is not an evaluator.

Drake makes that point better than I can, and by accident. He kicked off a build at the start of the talk, came back at the end, and found it still going after thirteen minutes, looking busy the whole time.

Position on a board tells you where a thing is. It does not tell you whether it is right, and the prettier the board, the longer you will believe it.