AMAP-ML/SkillClaw

There is a quiet elegance to what this framework attempts. Rather than hardcoding capabilities into individual agents, it lets skills emerge and improve through a collective evolutionary loop, where agents effectively teach and refine one another over repeated cycles. The agentic evolver acts as an orchestrator that selects, mutates, and propagates useful behaviors across the population, borrowing ideas from evolutionary computation and applying them to language-model-driven agents. The research framing is honest and the codebase reflects that. This is not a plug-and-play library you drop into a production pipeline. It is a sandbox for exploring whether agent collectives can bootstrap competence without exhaustive human curation of every skill definition. That is a genuinely interesting question, and the implementation gives you a working starting point rather than just a paper. The honest reservation is that the gap between the experimental results and anything deployable at scale remains wide. Documentation is sparse and assumes familiarity with multi-agent research conventions. -> Best for: ML researchers prototyping emergent skill acquisition in multi-agent environments.