Yes, but I don't think it's about the present, necessarily.

AI is still shit. There are prompt networks, what some people call agents, but presently models are still primarily trained as a singular models, not made to operate as agents in different contexts with RL on each agent being used to improve the whole indirectly.

Tokens will eventually become cheaper to the point where it will be possible to actually train proper agents. So we probably will end up with very powerful systems in time, systems that might actually be at least some kind of AGI-lite. I don't think is far off. At most a decade.