the article conflates two different things: what's easy for an LLM to write vs. what's safe to run. python wins on the first axis (smaller context, more training data, faster iteration). go wins on the second (compile errors caught early, explicit error handling). for a long-running autonomous agent that can't be supervised, you probably want the second. for a human-in-the-loop workflow where you can just rerun, the first.