"Our key insight is that LMs require carefully designed agent-computer interfaces (similar to how humans like good UI design). E.g. When the LM messes up indentation, our editor prevents it and gives feedback."