The SkillBrowser preprint, submitted on October 16 by Jin Ye, Yifan Jiang and Lan-zhe Guo, studies how much procedural content an agent sees. Showing only metadata for a growing catalogue reduces content per skill, but does not necessarily remove context pollution. Conversely, a single initial retrieval can miss needs discovered later.
The framework uses a meta-skill to identify the current capability need, formulate a targeted query and load instructions on demand. The authors report higher aggregate accuracy across six SRA-Bench subsets and five models compared with static retrieval and model-based selection. This concerns their experimental environment, not an automatic upgrade to every agent using skills.
The paper preserves two practical limits: some useful skills are not invoked and the process adds latency. Shrinking the visible catalogue is therefore not an end in itself; the agent must find the right procedure at an acceptable selection cost. The result informs harness design without establishing that one architecture dominates all tasks or that loaded instructions are safe.