AWS has published a guide to Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI, a method that trains models using the result of a multi-step task rather than labels for individual actions. In the described setup, SageMaker AI sends tasks to an agent hosted on Amazon Bedrock AgentCore, or on infrastructure such as Amazon EKS, Amazon EC2, or AWS Fargate through an AWS Lambda forwarder. The agent interacts with an environment over multiple turns and returns a reward when the task ends.

The guide uses Wordle as its example. The agent can make up to six guesses, receives feedback after each guess, and is scored according to the final outcome. AWS says the example uses Group Relative Policy Optimization to update a Low-Rank Adaptation adapter on the policy model. The dataset contains secret words, while the agent includes the environment and reward function.

According to AWS, a 100-step run using GPT-OSS-20B improved its solve rate on held-out words from 61% to 84% at the model’s default reasoning effort, and from 15% to 69% at the low effort used for training. AWS reports approximately $339 in training-token costs for that run, plus about $79 for the AgentCore runtime, based on the prices and usage described in the guide.

The post also covers deployment, IAM roles, dataset preparation, MLflow monitoring, evaluation, and supported models and Regions. AWS says the datasets, agent, and trained model remain in the customer’s AWS account and are not used to train base models.