SkillAdam proposes tracking past fixes and making smaller edits to AI agent instructions

A post describing a Tencent paper says SkillAdam logs past fixes as it rewrites instruction files that teach AI agents how to do tasks. It makes smaller changes when results are mixed, aiming to avoid undoing fixes that worked. The post reports 28.3% average accuracy on long shopping and travel planning tasks, versus 21.7% for SkillOpt, while using about a third as many tokens.