sandboxing agent actions on M3 Max

[southpatch]

Thanks, that helps. I ended up in about the same place using PyTorch. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.