I built a three-agent pipeline for research tasks and spent more time debugging their interactions — message loops, conflicting assumptions, silent failures — than I ever spent on single-agent prompts. The demos of agent swarms look magical until you operate one. Is multi-agent coordination a genuinely promising paradigm, or are we adding complexity faster than capability? What patterns actually work in production?