Self-evolving agents accumulate skills by appending successful procedures and failure fixes, but this leads to redundant and bloated skill definitions. SkillZip, introduced by Hugging Face researchers, addresses this by compressing skills without requiring evaluation rollouts, which are costly and dependent on the evaluation set.
The method works by finding the shortest faithful structural explanation of a skill, following the principle of 'explain once, reference many.' Repeated rules are stated once at the appropriate scope, repeated action sequences are factored into shared procedures, and only differences are kept as explicit exceptions.
SkillZip formalizes this as a typed minimum description-length objective over a skill contract and a residual, with a hard coverage constraint for every extracted trigger, workflow edge, tool requirement, obligation, and output field. This ensures that rare exceptions are preserved by construction.
The method offers two modes: a one-shot mode with a single structured extraction call and deterministic optimization, and a continual Zip-on-Write mode that integrates self-evolution patches without replaying tasks or reparsing the full history. Experimental evaluations demonstrate its effectiveness in compression performance, generalizability, and cost overhead.