SkillZip addresses a context-window problem for agent systems that load reusable skills at inference time. As libraries grow, retrieving an entire skill package can waste context, while naive text compression can break the procedure the skill is supposed to preserve.
The paper proposes representing skills as contract-bearing execution graphs and compressing those graphs into smaller routines that remain executable and expandable. That lets an agent expose only the pieces needed for a task while keeping the procedural contract attached.
The work is aimed at developers building large agent skill libraries, not end users. Its main claim is architectural: reliable reuse needs smaller procedural units than whole documents, especially when agents must update and compose skills over time.