Do agent skills actually work? What the research says
SkillsBench tested agent skills across 86 tasks and 7 agent-model setups. Curated skills raised pass rates by 16.2 points; self-generated skills did not help.
Updated Nov 10, 2026 · 3 min read
SkillsBench (arXiv 2602.12670) is the first large benchmark built to measure whether agent skills help. It covers 86 tasks across 11 domains and tests 7 agent-model configurations over 7,308 runs.
- Curated skills raised the average pass rate by 16.2 percentage points.
- Skills the model generated for itself gave no benefit on average.
- Averages hide variation between tasks, so test skills on your own work.
What this means for you
Pick skills written and refined by people who do the task well, then measure. Popularity is a useful first filter, which is why Alyn ranks by stars, installs and freshness, but your own before/after test is the final word.
FAQ
Is a skill better than a bigger model?
They are separate levers. A good skill improves the same model on the same task; a bigger model raises the baseline. Test both on your own tasks.
Should I let the agent write its own skills?
Use the agent to draft, then edit hard. Unedited self-generated skills did not help on average.
Next guide
Are agent skills safe? How to check a skill before you install it