Holy shit...Airbnb just turned customer support into a self-improving AI lab 🤯 Their new paper, Agent-in-the-Loop (AITL), shows how embedding human feedback directly inside live support workflows creates a data flywheel that retrains the model every few weeks not months. Instead of offline annotation marathons, AITL collects 4 real-time feedback signals from human agents: • Which AI response they preferred • Why they chose it • Whether the retrieved info was relevant • What knowledge was missing Those signals continuously retrain retrieval, ranking, and generation models cutting iteration time and boosting performance: +11.7% retrieval recall +14.8% precision +8.4% helpfulness +4.5% agent adoption The result? A system that learns while it works. No more static models. No more months-long retraining cycles. This is how AI becomes truly adaptive humans in the loop → agents in the loop → infinite improvement. Read full paper: arxiv. org/abs/2510.06674