Great idea! Are you also considering fine-tuning models as a "side-effect"? Basically, you collect traces of the agent's mistakes from your proxy, and you fine-

Great idea! Are you also considering fine-tuning models as a "side-effect"? Basically, you collect traces of the agent's mistakes from your proxy, and you fine-tune a smaller model that you can offer specifically for that agent as a cheaper alternative?