Coding agents generally are built for a chat log. Working with them on a canvas is a different problem: they need to see what’s there, where it is, and what’s off-screen. @max__drake wrote up how we’ve been building the harness that gives them that context: …↗