This is pretty amazing. I had some of this functionality implemented in a half assed way for my agent (syncing state to R2), but this seems like a game changer and has all the features needed for a safe sandbox with persistent agent states.