A survey and public compendium tries to fix the fact that 'agent' has no standard definition for evaluation purposes

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.11018 (submitted 2026-09-10) organizes agent evaluation around five dimensions of agenticness: environmental interaction, learning and adaptation, autonomy, goal-directed behavior and temporal coherence, synthesizing the metrics and benchmarks used for each. It ships the Agent Compendium, a public resource at agent.duketrustlab.com that organizes and extends the evaluation methods found. Useful mainly as a map of where agent evaluation is thin, which the authors identify per dimension rather than asserting in general.