# SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

**Authors:** Xiangyi Li, Yimin Liu, Wenbo Chen, Bingran You, Zonglin Di, Yifeng He, Shenghan Zheng, Kyoung Whan Choe, Jiankai Sun, Shuyi Wang, Chujun Tao, Binxu Li, Xuandong Zhao, Hejia Geng, Xiaojun Wu, Junwei Zhou, Xiaokun Chen, Hanwen Xing, Yubo Li, Qunhong Zeng  
**arXiv:** [2602.12670](https://arxiv.org/abs/2602.12670)  
**Date:** 2026-02-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.12670v4/

## Key points

- SkillsBench measures the impact of reusable Agent Skills by comparing the same agents on the same tasks with and without Skills, isolating their contribution rather than reporting only overall performance.
- The benchmark covers 87 containerized tasks across eight domains, selected from 400 submissions by 142 contributors. Tasks include human-written instructions, fixed environments, reference solutions, and deterministic verifiers.
- Skills were curated separately from task solutions, and task instructions do not tell agents which Skills to use. This is intended to test reusable expertise rather than access to hidden answers.
- The evaluation tested 18 model–harness configurations across four terminal-agent harnesses under matched conditions. Each task-condition cell targeted three trials, with scores averaged across trials and tasks.
- Curated Skills improved average task-macro pass rates from 33.9% to 50.5%—a 16.6-point increase, or a 25.5% normalized gain. All 18 configurations improved, though gains ranged from 4.1 to 25.7 points.
- Results depend on the model–harness combination. OpenHands with GPT-5.5 achieved the highest Skills-enabled pass rate at 67.3%; OpenHands with GLM 5.1 had the largest gain at 25.7 points. The same model also performed differently across harnesses.
- Benefits varied by domain: Natural Science gained the most on average (28.8 points), followed by Media and Content Production (24.1) and Cybersecurity (18.9). Skills helped less in Mathematics and Operations Research (9.7) and Software Engineering (11.6).
- Skills can sometimes hurt performance. Thirteen tasks had negative average changes, often because Skills encouraged overly complex workflows, displaced better default approaches, or suggested brittle methods.
- Focused Skills performed best: Tasks with one to three Skills gained more than those with four or more, and compact or standard-length documentation outperformed comprehensive prose.
- Self-generated Skills underperformed: In all three tested configurations, they fell 8.1–11.5 points below the no-Skills baseline, while curated Skills improved those same configurations by 18.2–24.8 points.
- The findings are promising but have limitations. The benchmark uses high-quality curated Skills, terminal-based tasks, a limited set of models and harnesses, and deterministic verifiers.

## Summary

SkillsBench investigates whether Agent Skills—reusable packages of procedural instructions, scripts, and reference materials—help AI agents complete tasks, and how their effectiveness varies across models and agent harnesses. Unlike benchmarks that report only overall performance, SkillsBench isolates the contribution of Skills by comparing the same agents on the same tasks with and without them. The benchmark comprises 87 containerized tasks across eight domains, including software engineering, natural science, cybersecurity, finance, office work, and industrial systems. The tasks were selected from 400 submissions by 142 contributors. Each task includes a human-written instruction, a fixed environment, a reference solution, and a deterministic verifier. Skills were curated separately from task-specific solutions, and task instructions do not tell agents which Skills to use. This design aims to ensure that Skills provide reusable expertise rather than hidden answers. The authors evaluate 18 model–harness configurations across four terminal-agent harnesses. Each configuration is tested under matched no-Skills and curated-Skills conditions. Each task-condition cell targets three trials; scores are calculated by averaging pass rates across trials and then across tasks. The study also examines domain differences, Skill design and invocation, and a separate condition in which agents generate their own Skills. Curated Skills raise the average task-macro pass rate from 33.9% to 50.5%, a gain of 16.6 percentage points, or a 25.5% normalized gain. All 18 configurations improve, although the size of the lift varies substantially, from 4.1 to 25.7 points. OpenHands with GPT-5.5 achieves the highest pass rate with Skills, at 67.3%, while OpenHands with GLM 5.1 records the largest absolute gain, at 25.7 points. Harness choice also matters. Gemini 3.1 Pro reaches 60.8% with Skills in Gemini CLI, compared with 52.8% in OpenHands. Claude Opus 4.7 reaches 61.2% in Claude Code and 53.1% in OpenHands. The authors conclude that Skill efficacy depends on the entire model–harness system, not just on the Skill itself. Benefits vary by domain. Natural Science shows the largest average increase, at 28.8 points, followed by Media and Content Production at 24.1 points and Cybersecurity at 18.9 points. Mathematics and Operations Research gain 9.7 points, while Software Engineering gains 11.6. Skills appear especially valuable for tasks involving specialized procedures, data conventions, or precise output requirements. Skills can also reduce performance. Thirteen tasks show negative average deltas, often because a Skill recommends an unnecessarily complex workflow, displaces a better default approach, or leads the agent toward a brittle method it cannot debug. Design analyses find that tasks with one to three Skills gain more than those with four or more, and that compact or standard-length documentation outperforms comprehensive prose. Self-generated Skills underperform the no-Skills baseline in all three tested configurations, by 8.1 to 11.5 points. By contrast, curated Skills improve those same configurations by 18.2 to 24.8 points. Overall, the findings suggest that well-curated, focused Skills can substantially improve agent performance and sometimes help smaller models approach the unaugmented results of stronger models. However, the benchmark represents an optimistic setting: it uses high-quality curated Skills, containerized terminal tasks, a limited set of models and harnesses, and deterministic verifiers. The authors also note that additional context may contribute to the gains, so stronger length-matched controls are needed to distinguish procedural guidance from the effect of simply providing more text.