Amazon SageMaker AI adds multi-turn reinforcement learning

SageMaker AI introduces multi-turn reinforcement learning (RL), a serverless technique to fine-tune models for multi-step agent tasks. This feature trains models against custom agent environments by rewarding the complete sequence of decisions an agent makes, facilitating the specialization of smaller models to match larger ones on target workloads.

SageMaker's Multi-turn RL handles the full training loop, including rollout orchestration, trajectory collection, and checkpoint management, eliminating the need for custom infrastructure. Users can connect their agents running on various AWS services or custom infrastructure.

The offering includes built-in MLflow tracking to inspect agent trajectories, rewards, and traces. Evaluation jobs provide key metrics like reward, pass@k, and trajectory metrics for pre-deployment benchmarking.

Multi-turn RL operates as a fully serverless capability, meaning users only pay for tokens processed without needing to provision or manage infrastructure, making it a cost-effective solution.

Multi-turn RL is available now via SageMaker Studio and the SageMaker Python SDK. Specific model support varies by region, with models like Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B, and Gemma 31B listed for us-west-2 and us-east-1.