Lenny RachitskyHost of Lenny's Podcast | Author of Lenny's Newsletter

The multi-agent system trap

15 Jan 2026 · 1 min

Watch on YouTube

0:38

A short segment from a YouTube talk arguing that multi-agent systems are often oversold; splitting complex problems into many specialized agents and connecting them via ad-hoc protocols (e.g., gossip) rarely yields the expected ‘agent utopia’. The speaker highlights practical difficulties, especially for customer support applications, and expresses skepticism that current modeling approaches and architectures are appropriate for such use cases.

Chapters

People have this notion of, I have this incredibly complex problem. Now I'm going to break it down into, hey, you're this agent, take care of this. You're this agent, take care of this. And now if I somehow connect all of these agents, they think that the agent utopia, it's never the case. This notion of I'm going to divide the responsibilities based on functionality and somehow expect all of that to work together in some sort of gossip protocol that is extremely misunderstood. Doing this in, say, a customer support kind of use case is incredibly hard to control. What kind of agent is replying to your customer? I don't think current ways of building and current model are right there in terms of building those kind of applications.