SkillZip shrinks agent skill libraries without evaluation runs

Self-evolving agents accumulate procedures and fixes across their working lives, and the resulting libraries reportedly balloon to five or six times their original size as the same rules get restated across branches, warnings, and examples. A new arxiv preprint proposes SkillZip, a compression method that skips the usual approach of running tasks to test what is safe to cut.

The guiding line is 'explain once, reference many.' SkillZip treats a skill as a typed contract and looks for the shortest structural explanation that still covers every rule, including rare ones, with a hard coverage constraint the authors say gives a rare-rule preservation guarantee independent of task frequency. Two implementations are described, a one-shot mode using structured extraction plus deterministic optimization, and 'Zip-on-Write,' which keeps the library compact as an agent continues learning without replaying past tasks.

The reported numbers are the interesting part. SkillZip posts a 31.2% average compression rate against 9.2% for a baseline the authors call SkillReducer, average performance of 0.577 versus 0.544, and a runtime of 286 seconds against a comparison roughly 3.5 times slower. On continual runs, skills reportedly grow to 1.6-1.9x the seed length instead of 2.5-3.7x uncompressed. A cross-model transfer test on LiveMath retains 0.97 of performance against 0.91 for the same baseline.

A few things to hold lightly. This is a fresh preprint rather than a peer-reviewed result, and the head-to-head reporting focuses on one baseline rather than the wider landscape of evaluation-based compressors. The paper also does not describe how the typed-contract extraction behaves when it misclassifies a rule, which is the failure mode a production team would most care about.

For anyone running long-lived agent stacks where the skill store keeps expanding, an evaluation-free compressor that promises to preserve rare rules is a useful thing to test on your own library before committing to another cleanup pipeline.