The mistake is treating an LLM output like a finished deliverable. I had an agent build agents for me last week. It got the broad shape right and missed details that mattered. That is the real workflow now: Prompt → generate → inspect → tune → repeat.