The AWS Machine Learning Blog presents a hands-on guide to Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI. MTRL trains models based on the outcome of an entire multi-step task rather than on individually labeled steps.
The guide walks through the full pipeline using Wordle as the example task. Steps include deploying the agent and reward function on Amazon Bedrock AgentCore, preparing the dataset, running training, and evaluating the model at serving settings.
Why it matters
Training on whole-task outcomes targets scenarios where labeling each intermediate step is impractical, which is relevant for multi-step agent behaviors.
Who should care
Practitioners working with reinforcement learning, agents, and the SageMaker and Bedrock AgentCore ecosystem who want a reproducible example to follow.