Build efficient agents with knowledge distillation on Amazon SageMaker AI
Summary
- Amazon SageMaker AI now supports sequence-level knowledge distillation (SeqKD) to fine-tune efficient agentic models using teacher trajectories.
- SeqKD transfers behavior from a frontier teacher model via text responses alone, requiring no internal logit access or RL infrastructure.
- A walkthrough demonstrates generating shopping agent trajectories with a frontier model on Amazon Bedrock and filtering them using a programmatic scorer.
- Filtered trajectories are encoded into multi-step training rows to fine-tune a smaller student model using serverless supervised fine-tuning.
Why it matters
Knowledge distillation allows developers to replace expensive frontier models with smaller, cost-effective student models for multi-step agentic loops without sacrificing task quality.