SkillAxe boosts LLM-written agent skills by 28%, closing gap to human performance

The paper tackles a critical bottleneck in LLM agent frameworks: skill documents—structured natural-language instructions that guide agents—are notoriously hard for LLMs to write effectively. On the SkillsBench benchmark, human-authored skills improve pass rates by 16.2 percentage points, while LLM-authored skills provide no measurable gain at all. The researchers introduce SkillAxe, which decomposes skill quality into four interpretable dimensions: quality impact, trigger precision, instruction compliance with fault attribution, and solution-path coverage. From these, it generates structured improvement briefs that require no ground-truth labels, test suites, or environment rewards, allowing LLMs to self-refine iteratively.

SkillAxe proves itself as a continuous improvement engine. On SkillsBench, it improves pass rates by 28% relative over unimproved LLM skills and closes 47–67% of the gap to human-authored skills. Even more impressively, on SpreadsheetBench—a complex, real-world spreadsheet automation task—a skill library built with SkillAxe learned from past agent trajectories and boosted pass rate from 16.0% to 52.0% using only 22 skills. The results suggest that fully autonomous, self-improving agent systems are feasible without expensive human annotations or hand-crafted tests.