Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems

A research team has proposed Skill-MAS, a method for automatically generating multi-agent systems that separates experience retention from parametric model updates, according to a paper posted to arXiv on June 17, 2026 [1]. The approach addresses a persistent tension in automatic Multi-Agent Systems (MAS) generation driven by large language models. Inference-time MAS methods use frozen frontier LLMs but repeat identical searches without learning from past experience, while training-time MAS internalizes experience through gradient updates but is limited by the smaller models it can accommodate [1]. Skill-MAS introduces a third path by treating the high-level orchestration capability as an evolvable Meta-Skill, encoding strategy-level principles that span task decomposition, agent engineering, and workflow orchestration [3]. The Meta-Skill is refined through a closed optimization loop with two stages. Multi-Trajectory Rollout executes multiple independent rollouts per task under the current skill, converting single-trial outcomes into distributional statistics that separate genuine capability from execution stochasticity [4]. Selective Reflection then prioritizes the most volatile and difficult tasks using a joint uncertainty–difficulty score with adaptive elbow truncation, and applies hierarchical contrastive analysis — first within-task, then cross-task — to diagnose systemic failure modes and distill the evidence into generalizable principles [4]. Experiments across four complex benchmarks and four distinct LLMs showed what the authors describe as remarkable performance gains while maintaining a favorable cost-performance trade-off [1]. The evolved Meta-Skills also demonstrated robustness and transferability to unseen tasks and different LLMs [2]. The paper was submitted to the computer science subcategory of Multiagent Systems on arXiv, an open-access repository of electronic preprints that has hosted scientific papers since August 1991 and now receives roughly 24,000 submissions per month [10]. The work appears under arXivLabs, a framework launched by the repository to allow community collaborators to develop and share experimental tools directly on the site, guided by values of openness, community, excellence, and user data privacy [8].