ai-trains-ai: RL-training an AI agent to RL-train AI agents.

A developer built a meta-RL pipeline where an AI agent (Qwen3-35B with LoRA) writes complete reinforcement learning training jobs , including environments, rewards, and hyperparameters , then submits them to real GPUs on RunPod. The agent itself is then RL-trained by rewarding it when the inner models improve. Over 54 training steps, reward climbed from ~0.0 to ~0.63, with skills transferring to a held-out task family.