SkillZip: A New Framework for Compressing LLM Agent Skill Libraries

A recent paper published on arXiv (2608.05604) presents SkillZip, a framework aimed at effectively managing extensive skill libraries for Large Language Model (LLM) agents. As LLMs increasingly function as agents with procedural knowledge in reusable skill packages, the libraries' size expands, complicating the exposure of the minimal executable context within limited budgets. Current systems face difficulties in reusing routines below the whole-skill level, maintaining procedural contracts during compression, ensuring that compressed routines remain executable and expandable, and updating the library as skills change. These issues arise from a mismatch in units: skills are retrieved as packages, compressed as text, and transformed into execution graphs post-retrieval. SkillZip introduces an execution-aware procedural abstraction framework that allows for contract-preserving compression on section-level graphs. It rewrites recurring contract-valid motifs into reversible ported abstractions, facilitating efficient and dependable reuse of procedural knowledge. The paper was announced as a cross-type submission on arXiv.