SkillClaw — Making Multi-Agent Skills Evolve Collectively from Real-World Usage

[Upper9527]

Hi Hive community!

We recently released SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems. Since Hive focuses on production multi-agent orchestration with self-improving capabilities, we thought this community would find it relevant.

The problem

LLM agent skills remain static after deployment. Different users repeatedly rediscover the same failure modes and workarounds, but there's no mechanism to convert these distributed experiences into systematic skill improvements.

Our approach

SkillClaw continuously aggregates interaction trajectories from multiple users and processes them with an autonomous evolver that:

• Identifies recurring behavioral patterns across users
• Refines existing skills or creates new ones
• Synchronizes updated skills across all users via a shared repository

The whole process is transparent to users — they just use their agent as usual, and skills evolve in the background.

Results

• Significantly improved Qwen3-Max on WildClawBench
• Reached moved honeycomb to oss #1 Paper of the day on Hugging Face
• Supports multiple OpenClaw-style agent frameworks

Links

• GitHub: https://github.com/AMAP-ML/SkillClaw
• Paper: https://arxiv.org/abs/2604.08377
• HF Paper Page: https://huggingface.co/papers/2604.08377

Feedback and ideas for collaboration are very welcome!