A r/LangChain post argues most multi-agent LangGraph systems are architecture theater — orchestration overhead that doesn't beat a single agent with tools on eval.