How to design a first multi-agent system without wasting tokens
A newcomer asks how to structure a multi-agent setup so it works well without burning tokens. The practical answer emerging from the thread: one orchestrator plus a few narrow specialists, each with only the tools it needs, short summaries passed between agents instead of full transcripts, and a verification step that checks output against the original request before anything is sent or saved.