Advances in Agent Training and Skill Management

New research challenges assumptions about agent skills: they primarily stabilize execution (65.7% procedural anchoring) rather than inject knowledge, with precision dropping as skill pool grows. Latent On-Policy Self-Distillation (LOPD) achieves strong agentic tool use with <30% rollout budget. A massive dataset of 3.8M agent skills across 282K repos is now available for mining, providing a rich resource for skill discovery.