Sorry, that section of the documentation is out of date; python 3.8 is supported and tested. I’ll update that section in the docs, and also check whether we can run with 3.9.
Turns out python 3.9 will need a little more work - there’s nothing in ml-agents code specifically that prevents 3.9, but some of the dependencies don’t work out of the box.
Great it works with 3.8 now as well. I remember I tried setting up the ml-agent project with students when it was still using tensorflow 1.7 and it was such a pain to get it working on all the different laptops. So far it seems it goes a bit smoother now but I keep my fingers crossed.