SkillClaw — A Framework for Collective Skill Evolution (related to RFC #1865)

[Upper9527]

Hi DeerFlow community!

We noticed the exciting RFC #1865 on Skill Self-Evolution for DeerFlow — we're working on a closely related problem and wanted to share our approach.

We recently released SkillClaw, a framework for collective skill evolution in multi-user LLM agent ecosystems. While RFC #1865 focuses on single-agent skill self-evolution (an agent improving its own skills from its own experience), SkillClaw takes this one step further: it aggregates interaction trajectories from multiple users and uses an autonomous evolver to distill them into shared skill updates that benefit the entire agent cluster.

Core idea

• Skills remain static after deployment → similar failures are rediscovered by different users repeatedly
• SkillClaw continuously collects cross-user trajectories and processes them with an autonomous evolver
• The evolver identifies recurring behavioral patterns and translates them into skill updates (refining existing skills or creating new ones)
• Updated skills are synchronized across all users via a shared repository

Architecture

1. Client Proxy — intercepts agent requests, records sessions, syncs skills
2. Workflow Evolve Server — 3-stage LLM pipeline (Summarize → Aggregate → Execute)
3. Agent Evolve Server — autonomous agent-driven alternative using OpenClaw with full tool access

Results

• Significant improvement on WildClawBench with Qwen3-Max
• .. #1 Paper of the day on Hugging Face

Links

• GitHub: https://github.com/AMAP-ML/SkillClaw
• Paper: https://arxiv.org/abs/2604.08377
• RFC RFC: Skill Self-Evolution — Agent-Managed Skill Creation & Improvement #1865: RFC: Skill Self-Evolution — Agent-Managed Skill Creation & Improvement #1865

We believe SkillClaw's multi-user evolution approach could complement DeerFlow's single-agent skill self-evolution design. Happy to discuss!

[hetaoBackend]

Thanks for sharing this. The overall idea makes a lot of sense to us and is highly aligned with some of the directions we’ve been thinking about for DeerFlow.

In particular, it fits very well with our longer-term thinking around multi-tenant and multi-user scenarios. We’ve also previously looked into approaches like OpenSpace, so it’s great to see this direction being explored in a concrete way.

We’d be happy to stay in touch and exchange ideas as we continue exploring this area.