Ask2Skill: Learning to Seek Expert Help for Accumulating Reusable Agent Skills

Learning reusable skills from interactive experience lets large language model (LLM) agents improve continually, but learning from autonomous experience is bounded by the agent's current capability, since on tasks beyond it the agent fails and yields no successful trajectory from which new skills can be distilled. We propose Ask2Skill, an expert-assisted skill learning framework in which the agent consults an expert at capability gaps and distills reusable skills from its own guided execution. Ask2Skill addresses how the agent should seek help and how the help it receives becomes reusable skills. First, it treats help-seeking as a learnable meta-skill informed by a textual record of its capability boundary. Whether a request is necessary depends on how likely the agent is to succeed on its own, so we introduce consultation-local credit assignment, which measures the necessity of each request by the agent's autonomous success rate from the same state and optimizes the meta-skill according to this necessity. Second, it writes successful assisted trajectories into the skill repository and, when an autonomous branch continued from the state before the request fails, contrasts the two executions to extract what turned failure into success. On ALFWorld and WebShop, Ask2Skill improves success rate over the strongest baseline by 3.8 and 4.2 points. Further analyses show that reliance on the expert declines as skills accumulate and that the skills transfer to other backbones.