I've been building and using this way of spatially brain-dumping, refining, and prompting agents, and honestly enjoying it.

I kept running into the same small friction every day: I'd have an idea for a coding agent, but phrasing it so the agent didn't go sideways usually meant juggling Excalidraw and pulling context from a dozen places — a doc here, an issue there, etc. Bonsai is the native Mac app I built to close that gap. The board isn't just UI. It's a live structured graph, no parent-pointer tree, every relationship is an edge between nodes, so the agent can read it the same way you do. That's what lets you actually refine