@nr4 Sure thing! I think you did the right call there and I doubt using 100x more would have meaningfully/aesthetically added something drastic... But the sheer compute power available on current GPUs is still mindboggling to me sometimes...
One thing we did in my workshops back then was to record the agent traces as grayscale image over N frames, then slowly interpolate the parameters towards a target config and rerun the sim, rinse and repeat... We then loaded the images into a voxel volume and extracted a surface mesh for 3D printing... Very similar approach to my Type&Form project (only there I used reaction diffusion):
https://www.flickr.com/photos/toxi/albums/72157604724789091/