Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
arXiv·medium signal
This paper (2607.01084, published 2026-07-01) probes how brittle tool-using agents become when moved from static training distributions into open-world conditions, exposing failure modes in generalization. A cautionary, builder-relevant result for anyone shipping tool-calling agents beyond their training set.