Scoping Persistent Skill Edits to the Task Family That Earned Them Turns Six of Eight Harmful Deployments Into None
arXiv·low signal
arXiv 2609.29144 studies frozen-model agents that edit their own prompts and skills across a 12-round code-repair stream. Global skill memory scored 0.713, below the static agent's 0.775, because edits that helped one task family broke others. Retrieving each accepted skill only for the family that earned it raised utility to 0.816 and accepted 63 updates with zero harmful ones. The takeaway for self-improving agents is to tag learned skills with the scope they were validated on.