My experimentation with Verblets also concluded plain functions are the most logical harness for LLMs.