LLM agents are increasingly self‑improving by writing and reusing textual skills, which are stored either in a single global document or as a flat pool of per‑task entries. Most evidence for these approaches comes from domains with structurally similar tasks. On long‑horizon workloads where each task requires a distinct solution, both forms break down in opposite ways: the document collapses into a generic discipline, while the pool inflates and its entries remain bound to the instance that created them. We argue that the missing unit of reuse is the solving procedure shared by a cluster of related tasks, and we build SkillGLoW (Global‑Local Weave) around this insight. Local skills generated during a task’s execution are aggregated into procedural families and compressed into de‑instantiated global priors; the instance‑specific details are regenerated per task rather than stored. A commit gate admits a prior only when real execution shows it does not degrade the deployed library.
Across four benchmarks (mathematical reasoning, terminal automation, software repair, and embodied control) and three models, the priors gain 17.2 points (hard) over the no‑skill baseline on average, with positive gains in all 12 continual‑improvement runs, and 18.0 points with local regeneration. The library holds one prior per procedural family, achieving a 3.6× compression compared to the per‑task pool. Under the same protocol, GLoW outperforms a published single‑document optimizer on 15 of 21 cells. Unmodified, the library lifts success on unseen ALFWorld tasks from 73.9% to 83.9%, evidence that what transfers is procedure rather than task memory.
Review