When LLMs write code to accomplish a task, that code has to actually run somewhere. And right now, the options aren't great. Spin up a sandboxed container and you're paying a full second of cold start overhead plus the complexity of another service. ...