# How to train a model on multi-turn tasks with reinforcement learning on Amazon SageMaker AI

> Original title: Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide

- Source: AWS Machine Learning Blog
- Published: 2026-10-23T17:48:54.000Z
- Bad Influence: https://badinfluence.news/items/cqut8va1b5d589xuqjy6z3bnt
- Original: https://aws.amazon.com/blogs/machine-learning/multi-turn-reinforcement-learning-on-amazon-sagemaker-ai-a-hands-on-guide/

## Summary

AWS shows how to use Multi-Turn Reinforcement Learning on Amazon SageMaker AI to train a model on episode-level outcomes, using Wordle as an example. The post covers deploying the agent to Amazon Bedrock AgentCore, preparing the dataset, running training, and evaluating at serving settings. In its run, gpt-oss-20b rose from a 61 to an 84 percent solve rate on held-out words at default reasoning effort, for about $339 in training tokens.

## Why it matters

The post walks through deploying an agent, preparing a dataset, and training with an outcome-based reward, including the reward-design mistakes it hit along the way. It is useful for teams building multi-step agents.