If you just pipe an LLM straight into a feedback loop with a console, it is going to destroy things rather soon, depending on the context window and other factors. All the agentic fluff is basically the workaround for LLMs going off the rails.