“The recursive language model (RLM) harness provides a model with access to a variable in a persistent Python REPL, allowing it to manage its context and call sub-LLMs.”