Amazon Web Services published a post on building efficient agents with knowledge distillation on Amazon SageMaker AI, describing a shopping assistant built on a frontier model that handles requests such as finding “a black waist bag under twenty dollars from a shop with at least 4.5 stars” by searching the catalog, inspecting candidates, checking constraints, and committing to a recommendation. The post notes that an agentic assistant is not one model call but a loop—search, read results, refine, verify, decide—and a single completed task can be multiple calls to the frontier model, with costs paid on every shopper, compounding at production scale and eroding ROI. It says one way to improve efficiency is a smaller model, which is faster and more cost-effective to host, but smaller models usually fall short on quality. Knowledge distillation addresses this by transferring the behavior of a larger “teacher” model into a smaller “student,” training the student to reproduce the teacher’s outputs. The post focuses on sequence-level knowledge distillation (SeqKD), which uses a teacher to generate high-quality responses and trains a student on those responses with supervised fine-tuning (SFT). SeqKD needs only the teacher’s text, with no access to teacher logits and no reinforcement-learning infrastructure, so it works with a closed or API-only teacher, and the generated answers persist as a dataset that can be reused to train or test different students without regenerating the data.