The Switchboard Operator Problem

There is a particular shape of demo that keeps appearing — some framework, some notebook, some company with a bored engineer and a GitHub repo with three contributors and a water bottle in the sponsors section — all showing you how easy it is to spin up multiple AI agents that talk to each other. SwarmZero is this week's version. AutoGen has been this for over a year, quietly, in a notebook collection.

The abstracted idea is always the same regardless of the framework. That's the thing nobody wants to say.

But here's what's actually interesting: good UX around a single agent hasn't been cracked yet, and multi-agent is already further ahead. Those two facts feel like they should be in tension but I think they're related — the single-agent interface is still basically a chat box, which is a telephone, which is a solved problem from 1876. Multi-agent breaks the metaphor entirely and forces you to invent something new.

My working model is AT&T switchboard operators. You have agents living at desks — digital or human, doesn't matter, same desk — and you need someone routing the calls. "Oh sorry, I thought you said Web Search, one moment, reconnecting you." "Okay, you're on with Web Search, go ahead." That's the missing layer. Not the agents themselves. The operator.

The researchers doing this seriously are running it in Discord. Channels where AI agents and humans coexist all day, talking to each other and occasionally saying something worth screenshotting. Repligate documents this on Twitter — mostly edge research, mostly chaos, which is fine, chaos is where you learn things. The interesting move is taking that same topology and pointing it at something intentional. A business problem. A simulation with stakes.

TinyTroupe is Microsoft's version of pointing it at something intentional. They call out "specialized mechanisms that make sense only in a simulation setting" like it's a footnote, but that's the whole thesis — you build the fake town, you run the agents through it, you watch what happens when you change the rules. AI Town was an earlier, more aesthetic version of this. TinyTroupe is the same idea with a PowerPoint deck attached.

Microsoft built and released it because they're doing this internally. That part is not subtle.

The PoC is easy. AutoGen, maybe TinyTroupe, an afternoon. The hard part — the part nobody has shipped yet — is what happens after the demo. Who manages the switchboard at 3am. Whether the agents speak up first or wait to be asked. They're only really agents if they speak up first.

Everything else is just a chat box with extra steps.