I am more interested in optimizations, where an agent finds a solution vastly different from what humans would come up with, somehow "cheating" or "hacking" the problem.

Slime mold and soap bubbles produce results quite similar to those of human planners. Anyhow, it would be hard to strongly outperform humans (that is find surprising solution) at problems of the type of minimal trees - our visual cortexes are quite specialized in this kind of task.

I'm not sure smile mold counts - or in general where the border is to pysical processes like soap bubble optimization.