Unity-Technologies/ml-agents — reverse-engineered prompt

Build me a Unity toolkit that lets game developers turn their game scenes into training playgrounds for intelligent agents. I want Unity users to add agents to 2D, 3D, VR, or AR scenes, define rewards and observations, then train those agents from Python using reinforcement learning or learning from demonstrations.

Please include example environments, support for one agent, teams that cooperate, and agents that compete against each other. Let people run curriculum learning, randomize environments for stronger training, and train with several Unity instances at the same time. Trained agents should be able to run back inside Unity for NPC behavior, automated testing, or trying out game design ideas.

Also include a simple Python API for researchers, ways to connect environments to Gym and PettingZoo style workflows, and a clean path for adding custom training methods. Add useful docs and setup instructions so a hobbyist can get an example training run working. Look up the current Unity package docs online if you need to.