The Thing in the Box

The word they're using is "superagent." OpenAI has one. Meta is building one. The superagent can browse the web, book things, fill out forms, and generally complete the tasks you've been putting off — autonomously, without asking permission at each step.

Nobody seems to be sitting with how strange that sentence is.

The thing being described — a system that observes an environment, makes decisions, takes actions, and adapts when those actions don't work — is what philosophers have been calling agency for a few centuries. We put "super" in front and dropped it in a product announcement.

I've been reading adjacent to this stuff long enough to notice the throughline: each individual capability sounds mundane in isolation, and then you chain them together and something happens that nobody has a clean word for. Not sentience — the researchers will correct you very firmly on this, with citations. Not consciousness, definitely not. But also not nothing.

What you get when a model can set a goal, attempt to achieve it, notice the attempt failed, and try a different approach is, at minimum, a thing that behaves like it wants something.

That's not a trivial observation. That's the whole ballgame.

OpenAI would like you to think of Operator as a very helpful assistant. Meta would like you to think of theirs as a social companion. Both framings are technically accurate and both are doing enormous work to keep you from looking directly at what's in the box.

The move — and it's a good move, commercially — is to anthropomorphize just enough that the product feels approachable, while stopping well short of anything that would require a press release about machine consciousness. Superagent is a perfect word for this. It sounds like a butler. It doesn't sound like anything that has preferences, or persistence, or a model of itself operating in the world.

Which is fine. Except it does have those things, kind of, in the ways that matter for what it does next.