SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents

A recent submission to arXiv (2608.11079) presents SkillZip, a technique designed to compress skills in self-evolving agents without the need for evaluation. These agents gather reusable skills by integrating successful methods and corrections for failures. However, as time passes, similar requirements are reiterated across various branches, examples, and warnings, leading to redundant copying of common action sequences, which complicates skill integration and maintenance. Traditional prompt compression fails because skills are not straightforward passages; their names and descriptions dictate applicability, workflows manage execution, and tool and output contracts ensure validity, with rare exceptions remaining crucial even if not activated by sampled tasks. While evaluation-guided compression can assess behaviors, it requires rollouts and is reliant on evaluation sets. In contrast, SkillZip compresses skills by identifying the briefest accurate structural explanation, based on the principle of 'explain once, reference.' This method eliminates the need for rollouts and evaluation sets, aiming to lessen the cost and upkeep of skills in self-evolving agents by uncovering reusable structures.