Preregistration for Experiments with AI Agents

The paper advocates for the adoption of preregistration practices in experiments involving AI agents, emphasizing the need for methodological rigor as these agents increasingly engage in decision-making roles. It identifies various researcher degrees of freedom, such as model selection and prompt wording, that can compromise experimental integrity. By proposing a tailored preregistration template, the authors aim to enhance the credibility of research in this rapidly evolving field, which is crucial for practitioners developing and deploying AI agents in real-world applications.