AWS has published a Wordle walkthrough for training agents over several interacting steps on SageMaker AI. The training service and game environment operate separately, with the secret answer hidden from the model. The example details separately billed token and runtime costs and shows how an unsuitable reward design can favor unsuccessful games.
AWS has published a SageMaker AI walkthrough for reinforcement learning across multiple interacting steps, a method that trains behavior through rewards. The Wordle example trains an agent over complete tasks. SageMaker handles training and the policy model; AgentCore hosts the game and reward function. The model receives letter feedback over up to six guesses, while only the environment and scoring function see the secret answer.[1], [2]
The dataset supplies tasks rather than correct answers or demonstration trajectories. It uses 640 training words and a separate 128-word validation set. AWS describes a 100-step run with 105,486 rollouts. Token charges came to 338.50 dollars at the stated Northern Virginia list prices, with approximately 79 dollars more for AgentCore processor and memory use. Longer trajectories and more verbose reasoning increase consumption.[1]
An initial reward design awarded positive scores to lost games and sometimes favored slower solutions. The revised design gives the outcome the dominant reward and bounds the smaller shaping component. The accompanying code addresses authentication, concurrency and failed model calls. AWS also distinguishes training validation from evaluation at serving settings, which use different decoding choices, and leaves additional training steps untested.[1]
