Lifecycle Governance of Agent Skills: From Collection and Recommendation to Evolution
Route skills just in time, learn from task execution, and evolve reusable skill libraries through attribution-grounded feedback.
Powered by MemTensor
Agent skills are becoming a reusable execution layer for coding agents, research agents, and workflow agents. SkillsVote starts from this large-scale setting: we have discovered over 🔥 1.68M SKILL.md files from open-source GitHub repositories, including over 💎 790K format-valid skills verified with the official Anthropic skill validator, making SkillsVote the world's largest open agent skill library 🌍.
At this scale, skill management is no longer about manually maintaining a short curated list. Agents face three linked problems: which skills to load before a task, how to tell whether a skill actually helped during execution, and how to update the library without accumulating noisy or unverified experience.
SkillsVote treats skills as lifecycle-managed artifacts. It connects collection, profiling, just-in-time recommendation, trajectory-based attribution, and feedback-driven evolution into one loop:
- Collect and profile skills from open-source or private skill libraries.
- Recommend relevant skills before task execution, instead of loading a large static skill list.
- Attribute task outcomes after execution using trajectories, skill usage, and verifier signals.
- Evolve the skill library by updating or creating reusable skills from grounded, attributed feedback.
- 📄 [2026-05-19] Technical Report. We released the SkillsVote technical report on arXiv: arXiv:2605.18401.
- 🧭 [2026-05-18] Local Skill Integration. We released the first version of
skills-vote-local, enabling local/private skill recommendation with configurable retrieval strategies. - 🌅 [2026-04-09] Special Share. Core contributor's share on Linux.do.
- 📣 [2026-04-08] Social Launch. Our launch announcement is now live on WeChat Blog and rednote.
- 🚀 [2026-04-03] Launch Day! Published the very first open-source release of our recommendation and evaluation demos.
SkillsVote is being open-sourced in stages to support transparent research on agent skill collection, recommendation, attribution, and evolution.
- Skill profiling and preprocessing. Analyze skill runtime requirements, dependencies, quality, and verifiability.
- Benchmark evaluation scripts. Release scripts and configs for reproducing the main experiments reported in our paper.
- Hosted SkillsVote skill integration. Release the
skills-voteagent skill that connects agents to the hosted SkillsVote service for cloud-based recommendation and attribution-grounded feedback. - Local SkillsVote recommendation integration. Release
skills-vote-localwith configurable local/private skill recommendation strategies, including agentic search and vector search. - Local SkillsVote attribution and evolution integration. Extend
skills-vote-localwith attribution-grounded feedback and local skill library evolution. - Main experiment trajectories and results. Release benchmark trajectories and aggregated results to support inspection and reproduction of the reported experiments.
SkillsVote is evaluated on agentic coding and terminal challenge benchmarks, including Terminal-Bench Pro, Terminal-Bench 2.0, and SWE-Bench Pro.
The results show that just-in-time skill recommendation and feedback-driven evolution improve agent performance on long-horizon tasks. Detailed reproduction instructions, benchmark setup configs, and scripts are documented in docs/experiment.md.
| Integration | Best for | Requires |
|---|---|---|
skills-vote |
Using the hosted SkillsVote service for cloud-based skill recommendation and attribution-grounded feedback. | SKILLS_VOTE_API_KEY |
skills-vote-local |
Recommending skills from a local or private SKILL.md library without relying on the hosted index. |
Local config; no SkillsVote API key for agentic search |
Use this integration when you want agents to retrieve skills from the hosted SkillsVote service and submit post-task feedback for attribution.
Supercharge your agents (Codex, Claude Code, OpenClaw) by integrating SkillsVote directly! Just drop this prompt into your agent:
Install the `skills-vote` skill following https://raw.githubusercontent.com/MemTensor/skills-vote/main/integration/skills/INSTALL.md
Use the following values:
- `SKILLS_VOTE_API_KEY`: "YOUR_API_KEY"
- `GH_TOKEN`: "YOUR_GITHUB_TOKEN"Are you a CLI warrior? Set it up manually based on your OS:
Windows PowerShell
[Environment]::SetEnvironmentVariable("SKILLS_VOTE_API_KEY", "YOUR_API_KEY", "User")
npx skills add MemTensor/skills-vote --skill skills-voteMacOS/linux (Bash/Zsh)
# For zsh, use ~/.zshrc instead
echo 'export SKILLS_VOTE_API_KEY="YOUR_API_KEY"' >> ~/.bashrc && source ~/.bashrc
npx skills add MemTensor/skills-vote --skill skills-voteNote
Don't forget to replace YOUR_API_KEY with your actual key!
Use this integration when your skills are stored in a local or private SKILL.md library and you want recommendation without the hosted index.
Install the `skills-vote-local` skill following https://raw.githubusercontent.com/MemTensor/skills-vote/main/integration/skills/INSTALL.mdnpx skills add MemTensor/skills-vote --skill skills-vote-localAfter installation, open the installed skill root and configure configs/config.yaml. See Install SkillsVote Skills for the full configuration flow.
SkillsVote builds on the broader agent skill and agentic benchmark ecosystem. We thank the maintainers and contributors of Anthropic Skills, Harbor, and open-source agent skill repositories for making this research possible.
If you find SkillsVote useful for your research or development, please cite:
@misc{liu2026skillsvotelifecyclegovernanceagent,
title={SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution},
author={Hongyi Liu and Haoyan Yang and Tao Jiang and Bo Tang and Feiyu Xiong and Zhiyu Li},
year={2026},
eprint={2605.18401},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2605.18401},
}This repository is licensed under the MIT License. See LICENSE.