Researchers have proposed a new approach to evaluating AI agents, addressing issues with outcome finality and cross-unit separation in current evaluation methods. Outcome finality refers to the condition where an agent's actions are considered complete, while cross-unit separation ensures that each run is treated as an independent trial. The authors argue that these conditions are not necessarily met at the end of a stopped run, which can lead to inaccurate scores. They introduce a completion argument and propose an open-effects record to track operations or resources that may still affect the outcome after the endpoint. --- Why it matters: This matters because current evaluation methods can produce misleading results, affecting the development and deployment of AI agents. By addressing these issues, researchers can improve the accuracy of agent evaluations, leading to better decision-making in fields like robotics, finance, and healthcare. Source: https://arxiv.org/abs/2608.14940