RL-trained Qwen3.6-35B-A3B agent trains smaller models

The open-source agent writes full RL training jobs (environment, reward, dataset, hyperparameters) and submits them to real GPUs. It achieved higher scores on tasks it trained for.