// Evolving Meta-Skill for Multi-Agent Systems // Can a multi-agent system get better at orchestration without touching a single weight? Aut…

// Evolving Meta-Skill for Multi-Agent Systems // Can a multi-agent system get better at orchestration without touching a single weight? Automatic MAS generation has been stuck between two bad options. Inference-time methods use frozen frontier models but never learn from past runs. Training-time methods learn but are capped by small-model capability. Skill-MAS takes a third path. It treats the orchestration capability as an evolvable Meta-Skill, refined through a closed loop of multi-trajectory rollout and selective reflection that distills experience into strategy-level principles rather than memorized traces. Across four benchmarks and four different LLMs, the evolved Meta-Skills transfer to unseen tasks and other models, because the know-how lives in text at the strategy level, not in any one model's weights. You keep the frontier model and still accumulate experience. Paper: https://arxiv.org/abs/2606.18837 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Source: DAIR.AI (X) | 2026-06-20