“GPT agents” are cool experiments, but the demos look sort of fake?

By fake, do you mean something like "giving an exaggerated impression of how likely they are to be successful in their aims when the task is actually difficult, or in how impressive we ought to find what they can currently achieve?"

If so, I think that can lead us into a mistaken view, which would be to compare AutoGPT to The Arrival of the Mail Train. The story is that the Lumiere brothers screened a continuous 50 second still shot of a train coming straight toward the camera for an 1895 audience, and that they ran screaming out of the movie theater because they weren't sophisticated enough about film to tell the difference between the moving image of a train from the real thing.

In this analogy, our reaction to AutoGPT appears as that of the audience to The Arrival of the Mail Train, where we lack the sophistication to tell the behavior of ChaosGPT apart from that of a potentially effective world-destroying agent. It behaves and "talks" as if that's what it was doing, but probably in its current form has effectively zero chance of being even slightly successful in executing its goal. We drastically overestimate its chances.

But of course, AutoGPT is not like The Arrival of the Mail Train, because with film, we know that the apparent danger is harmless. By contrast, with AutoGPT, we have good reason to think the mechanism is only harmless for now. A better analogy is the Mogwai from Gremlins: a cute little pet, totally harmless for now, but which can become extremely dangerous unless you consistently comply with a set of strict rules to keep it performing safely, which either you or somebody else won't.