A new paper introduces SkillForge, a reinforcement-learning method for managing agent skill libraries through a fitness-driven lifecycle of trial, active, stable, and retired states. The approach uses a pre-RL evaluation phase to filter out low-quality skills before supervised fine-tuning, then continues managing skills during training through selective retirement, stabilization, and mutation operations to prevent obsolete or harmful skills from accumulating and degrading agent behavior. The authors report improved performance on agent benchmarks and release SkillFurnace, a dataset of annotated skill libraries and retirement events to support further research on skill management.
