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

This article announces Amazon SageMaker AI's new multi-turn reinforcement learning capability for fine-tuning AI agent models on complex, multi-step tasks.

• Multi-turn RL trains models against custom agent environments with full sequence rewards
• Enables smaller, lower-cost models to match larger general-purpose models on target workloads
• Fully serverless with no infrastructure provisioning required; pay only for tokens processed
• Integrates with Amazon Bedrock AgentCore Runtime, EKS, EC2, Fargate, or custom frameworks
• SageMaker manages training loop, rollout orchestration, trajectory collection, and checkpoint management
• Built-in MLflow tracking for inspecting agent trajectories, rewards, and traces
• Supported models: Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B, Gemma 31B in us-west-2

Multi-turn RL simplifies complex agent model training by eliminating custom infrastructure setup and providing managed end-to-end training capabilities.