arXiv:2603.15457v2 Announce Type: replace-cross Abstract: Tool-using artificial intelligence (AI) agents are increasingly adopted as systems that plan, observe their environment, and take actions over extended time periods. This shift challenges the current evaluation practice, in which models are tested in restricted, fully observable settings.
