Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide
AWS SageMaker AI now supports Multi-Turn RL to train agents on sequential task outcomes, not single prompts.
“MTRL took gpt-oss-20b from a 61 to an 84 percent solve rate on held-out words at its default reasoning effort, and from 15 to 69 percent at the low effort it trained at, for about $339 in training tokens.”
AWS has released a hands-on guide for Multi-Turn Reinforcement Learning (MTRL) on SageMaker AI, enabling practitioners to train small models on full agentic task sequences scored by end-to-end outcomes rather than step-by-step imitation. The pipeline integrates SageMaker with Amazon Bedrock AgentCore and demonstrates meaningful model improvement at low cost (~$339 in training tokens). This matters because it makes enterprise-grade agentic fine-tuning on private workflows accessible without exposing data to base model providers.