Skill-MAS
A research framework that evolves a single meta-skill file describing how an LLM should design and run a multi-agent system, refining it over rounds using contrastive reflection on past attempts.
Skill-MAS is the code release for a paper on evolving a meta-agent skill file that instructs an LLM to design and orchestrate a multi-agent system in three stages: task decomposition, agent engineering, and workflow orchestration into executable code.
An evolution round runs several trajectories per validation task with the current skill file, compares high- and low-scoring trajectories to synthesize improvement signals, then has an optimizer LLM rewrite the skill file using those signals; the best snapshot across rounds is kept. It supports multiple benchmark backends and is written in Python and Apache-2.0 licensed.