When people store files, they rely on folders nested inside other folders. An agent might correctly gather all notes from a meeting into one folder yet put that folder under the wrong project. Existing tests miss this mistake because they only check whether every file is saved, not whether the folder sits in the right place.
The researchers built a of 150 tasks where agents must organize scattered records into folder hierarchies. They tested 45 systems combining nine AI models with five different agent setups. A new Structure Score measures how well the hierarchy matches the original at every level, ignoring folder names.
The best system scored 83.2 out of 100 overall, but none exceeded 78 on meeting recaps or issue threads. Parent-level relations had higher error rates than leaf-level groups in every domain. This matters because misplacing whole folders undermines how people later find and use information.