SkillGLoW organizes agent skills into procedural families

Researchers have introduced a new method for organizing the knowledge that large language agents accumulate over time. Instead of storing every task separately or compressing everything into a single document, the system identifies the shared procedure across a group of similar tasks, retains only its skeleton, and regenerates the task-specific details on each new run.

The familiar problem: when an agent writes a textual skill after every task, the library either balloons into thousands of items each stuck in its original context, or gets compressed into a generic document that does not help with concrete tasks. SkillGLoW solves this in two stages, local and global. At the local level the agent writes a skill from the current run; at the global level, skills belonging to the same "procedural family" are merged into an abstract Prior, stripped of instance-specific detail — de-instantiated, in the paper's terminology. The missing details are regenerated in real time for each new task. A commit gate admits a Prior to the library only if a live run shows it does not degrade performance.

Experiments covered four domains — mathematical reasoning, terminal automation, software repair, and embodied control — across three base models. Against a no-skill baseline, the Priors added 17.2 points on average on hard tasks, with a positive gain in all 12 continual-improvement runs. Enabling local reconstruction as well pushed the gap to 18.0 points. The library remained lean: one Prior per procedural family, 3.6 times more compact than a flat per-task store.

Under the same evaluation protocol, SkillGLoW beat a previously published single-document optimizer in 15 of 21 comparison cells. The most convincing test came outside the training distribution: without any adaptation, the library lifted success on previously unseen ALFWorld tasks from 73.9% to 83.9%. That is strong evidence that what transfers is the solving procedure, not memory of the specific task.

The paper appears as a preprint on arXiv and has not yet undergone peer review. Code and data are available, enabling reproduction, but metrics were measured in simulated environments, not in production. It remains unclear how the method will behave when the number of procedural families scales to the thousands, or when the base model is replaced with an entirely new version.