Agent Skill Banks Overfit When Updated Locally — GSE Adds a Skill Relation Graph and Gains up to 180% Recall

arXiv 2608.06153·medium signalXBlueskyLinkedInCopy link

GSE (arXiv 2608.06153, Aug 6) attacks a failure mode in self-improving agents: treating skill evolution as a sequence of local updates produces overfitted skills that don't generalize. It maintains a Skill Relation Graph that explicitly models and co-evolves inter-skill relationships, adds cluster-based consolidation to abstract reusable capabilities, and uses replay-driven verification to catch behavioral regressions. Evaluated on bug-revealing test generation and false-positive bug report filtering across OpenHands and mini-SWE-agent, GSE improves precision 6.1-34.1% and recall 31.8-180.0% for test generation, with a 61.4% F1 gain on an internal industrial agent. Directly relevant if you maintain a growing skills directory that agents write back into.