Amazon SageMaker AI now has a documented workflow for training agents from the result of a complete multi-step task rather than from a correct answer at every turn. Multi-turn reinforcement learning lets a model act repeatedly in an environment and receive one reward when the episode ends.
AWS demonstrates the pipeline with Wordle. Each guess depends on feedback from earlier guesses, making the game a simple stand-in for workflows such as handling a support ticket or completing an internal business process. In the reported run, training raised the held-out Wordle solve rate of gpt-oss-20b from 61% to 84% at its default reasoning setting.
The training service sends prompts through an AWS Lambda forwarder to an agent hosted on Bedrock AgentCore or other infrastructure. The agent calls the policy model, interacts with its environment and reports the final score. Developers define how success is measured rather than supplying one ideal sequence of actions.
Wordle is only a teaching example, and frontier models already play it well. Real deployments need reliable reward functions, controlled environments and evaluation at serving settings. A poorly specified reward can teach shortcuts rather than the intended process, so the reported game result should not be read as proof that the same gain will transfer to a business workflow.