I’ve been spending a lot of time lately experimenting with multi-agent workflows and on the surface, the capabilities look incredible. You tie an LLM to a couple of tools, tweak a prompt loop and watch it solve tasks in real time. But once you try to m…