SkillForge gives agent skills lifecycles

SkillForge tackles a neglected weakness in skill-augmented agents: libraries that grow indefinitely can preserve outdated or harmful procedures. Its strongest contribution is treating skills as maintainable, testable assets rather than permanent memories.

• –Pre-retirement filters weak skills before supervised fine-tuning
• –Online reinforcement learning combines retirement, stabilization, and LLM-guided mutation
• –Results improve over SkillRL by up to 7.8% while keeping libraries compact
• –SkillFurnace exposes trajectories, fitness histories, lifecycle snapshots, and human-labeled retirement causes
• –The approach makes skill governance measurable, though its benchmark gains still need validation beyond ALFWorld, WebShop, and search QA