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

Amazon SageMaker AI now offers multi-turn reinforcement learning (RL), a new serverless model customization technique for fine-tuning models on multi-step, agentic tasks. SageMaker AI model customization lets you adapt foundation models using techniques such as supervised fine-tuning, reinforcement learning from verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF), without the undifferentiated heavy lifting of building and operating your own training infrastructure. Multi-turn RL extends this by training models against your own agent environment and rewarding the full sequence of decisions an agent makes across a task, helping you specialize smaller, lower-cost models to match or exceed the task accuracy of larger general-purpose models on your target workload.

Training models that power agents to reliably complete multi-step tasks is complex and time-consuming, often requiring custom infrastructure that takes weeks to build. SageMaker's Multi-turn RL offering handles this for you. You can connect your agent running on Amazon Bedrock AgentCore Runtime for fully managed hosting, or on Amazon EKS, Amazon EC2, AWS Fargate, or any infrastructure using the framework of your choice. SageMaker AI manages the full training loop, from rollout orchestration and trajectory collection to training and checkpoint management. Built-in MLflow tracking lets you inspect agent trajectories, rewards, and traces. Evaluation jobs report reward, pass@k, and trajectory metrics so you can benchmark a model before deploying it to a SageMaker AI endpoint or Amazon Bedrock. Multi-turn RL runs as a fully serverless capability, so you pay only for the tokens processed, with no infrastructure to provision or manage.

Multi-turn RL is available today through SageMaker Studio and the SageMaker Python SDK as part of Amazon SageMaker AI model customization. Supported models include Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B and Gemma 31B in us-west-2, and Nova Lite 2.0, GPT-OSS-20B in us-east-1. To get started with multi-turn reinforcement learning in SageMaker AI, visit the Amazon SageMaker AI documentation.