[jayd16]

You could run hundreds of agents in parallel and arguably that counts.

[aroman]

No, it wouldn’t. The hours in question are human experience, not that of the agent.

[jayd16]

Why not? You have far more results to review.

If you run a model on slower hardware are you getting more experience? Surely its a factor of model output reviewed and not human time.

[aroman]

Right, it's about how much time the human spent - the time spent by the machine itself is irrelevant. As you rightly point out: that is why we measure programmer experience by wall clock time, not CPU time :)