Skill-MAS evolves meta-skills to automate multi-agent system design

Existing LLM-based multi-agent systems (MAS) face a trade-off: inference-time approaches use frozen frontier models but cannot learn from past experiences, while training-time methods internalize knowledge through gradient updates but are limited to smaller, less capable models. Skill-MAS, introduced by Hehai Lin, Qi Yang, and Chengwei Qin, offers a third path that separates experience retention from parametric updates. By conceptualizing high-level orchestration as an evolvable Meta-Skill, the system refines its architectural knowledge through a closed optimization loop.

The loop operates in two stages. First, Multi-Trajectory Rollout samples a behavioral distribution for each task under the current Meta-Skill. Second, Selective Reflection adaptively picks priority tasks and applies hierarchical contrastive analysis to distill systemic experience into generalizable, strategy-level principles. Extensive experiments across four complex benchmarks and four distinct LLMs show that Skill-MAS not only delivers remarkable performance gains but also maintains a favorable cost-performance trade-off. Further analysis confirms that evolved Meta-Skills are highly robust and transfer effectively to unseen tasks and different LLMs, making Skill-MAS a practical path toward scalable, learning-based multi-agent automation.