I think LLM-driven agentic flow is useful in some cases, but in many other cases, deterministic code would be indeed much safer and more reliable. In the ideal

I think LLM-driven agentic flow is useful in some cases, but in many other cases, deterministic code would be indeed much safer and more reliable. In the ideal world, people can build a proof-of-concept with LLM agentic flow quickly, and if it seems to work like expected, then they should use LLM to convert that agentic flow to deterministic code!