Score every skill edit by rolling the task back and re-running it: a 'rollback reward' lifts agent skills 3.3–6.7 points over hand-crafted pipelines

arXiv 2608.01678·medium signalXBlueskyLinkedInCopy link

Skill-α reframes skill generation as reinforcement learning over sequential edits rather than heuristic consolidation, decomposing skill construction into individually evaluable edits. The novel signal is a rollback reward that scores each modification by comparing downstream task execution using the original skill against the edited one — solving the core problem that skills have no natural correctness label, only downstream utility. It gains 3.3 percentage points on CL-Bench and 6.7 on tau2-bench over the strongest baseline with a GPT-4o worker, and code is public; the transferable idea for anyone maintaining CLAUDE.md or SKILL.md files is to A/B each edit against real task replays instead of accepting agent-proposed doc changes on inspection.