AI agents that teach themselves new skills eventually drown in their own notes, and researchers have a fix that does not require re-testing everything to check the cleanup did not break anything.
A new paper describes SkillZip, a method for compressing the skill libraries that self-evolving AI agents build up by appending working procedures and fixes for past failures. The problem: the same rule gets restated across multiple examples and warnings, and repeated action sequences get copied rather than reused, so the file balloons and slows the agent down. Instead of shrinking that text as one flat block, SkillZip breaks a skill into its parts, including what triggers it, its workflow, and its tool and output requirements, then states each repeated rule once, at the level where it applies, folding duplicated action sequences into a shared procedure while keeping rare exceptions intact. It formalizes this as a minimum-description-length problem with a hard coverage rule: every trigger, workflow step, tool requirement and output field from the original skill must still show up in the compressed version.
That coverage guarantee is the real pitch. Prior evaluation-guided compression checks whether a shrunk skill still works by re-running sample tasks against it, which costs compute and only catches what the test set happens to exercise. SkillZip skips the rollouts entirely, running either a one-time extraction pass or a continuous one that folds in each new patch as an agent updates itself, which matters more as agents pile up longer self-written histories while context budgets stay tight.
The paper claims SkillZip beats rival approaches on compression, generalization and cost, but the abstract does not say against what baselines or at what scale, so file that under preprint claim rather than settled result.