Been reading through this sub and noticing a split I can't quite resolve.
On one side, half the posts are about LangGraph supervisors, CrewAI crews, multi-agent orchestration patterns.Reads like multi-agent is the future and everyone's heading there.
On the other side, every production system I've actually seen up close is a single agent inside a workflow. The LLM decides what to do next, picks the tool, drives the reasoning. The surrounding code handles the mechanics, parsing tool calls, retries, error handling, knowing when to stop
So I'm trying to figure out which one is real. Two questions:
If you have actual multi-agent in production (multiple agents handing work to each other, not just one agent with tools), what's the topology and roughly how many agents in the graph?
If you tried multi-agent and went back to a single agentic workflow, what made you reverse? Costs, debugging, latency, just couldn't get it stable?