Amazon SageMaker AI launches multi-turn reinforcement learning for AI agent model customization

Amazon Web Services has launched multi-turn reinforcement learning (RL) capabilities within SageMaker AI, introducing a new serverless model customization technique designed to fine-tune AI models for complex, multi-step agent tasks. The service allows developers to train models against custom agent environments while rewarding complete decision sequences across tasks, enabling smaller, cost-effective models to match or exceed the performance of larger general-purpose models on specific workloads. SageMaker's multi-turn RL offering integrates with Amazon Bedrock AgentCore Runtime for managed hosting or can connect to custom infrastructure running on Amazon EKS, EC2, AWS Fargate, or external platforms. The service manages the entire training loop, including rollout orchestration, trajectory collection, training, and checkpoint management, while providing built-in MLflow tracking for inspecting agent trajectories, rewards, and traces. As a fully serverless capability, users pay only for tokens processed without needing to provision or manage underlying infrastructure. The service is now available through SageMaker Studio and the SageMaker Python SDK, supporting models including Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B, and Gemma 31B in the us-west region.