We've been trying new patterns for organizing large numbers of agents over long periods of time for very little cost. We recently were asked to make a demo for a customer and decided to ask the agents to build a RCT2 clone.
This seemed hard, because RCT2 requires a lot of understanding of geometric space, a lot of art assets and animations, and a lot of complex interaction patterns.
The agents built a reasonable clone in ~48 hours, with just a claude max and codex max subscription. I spent the next week playing it, because I love coaster sims, and tweaked more of the gameplay elements that the agents missed.