> Isn’t this what AGI is by design?

Well, the "G" in AGI is kinda important. These are specifically games/puzzles.

> they have to be retrained from scratch

Is that true? Didn't DeepMind already build plenty of agents that are generally good at most computer games without being retrained?