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

Skill-MAS introduces a novel framework that enhances multi-agent AI systems by evolving meta-skills through a closed optimization loop, achieving significant performance gains while maintaining cost efficiency across diverse LLMs and tasks.

Skill-MAS addresses a fundamental architectural challenge in LLM-based multi-agent systems: the tension between leveraging powerful frontier models and accumulating learned experience. Existing approaches create a false choice—inference-time systems use cutting-edge models but waste computational resources by repeating identical searches, while training-time systems learn from experience but hit capability ceilings with smaller models and don't scale to frontier LLMs. The research decouples these concerns by treating orchestration strategy as an evolvable meta-skill separate from parametric model updates. This separation enables the system to maintain frozen, high-capability models while still accumulating architectural knowledge. The methodology employs two complementary mechanisms: Multi-Trajectory Rollout generates diverse behavioral distributions for each task, creating rich signal diversity, while Selective Reflection applies hierarchical contrastive analysis to distill experience into generalizable principles rather than task-specific memorization. For the AI development community, this approach represents meaningful progress toward systems that are simultaneously capable and adaptive. The demonstrated transferability across unseen tasks and different LLMs indicates the evolved meta-skills capture generalizable orchestration principles rather than brittle patterns. The favorable cost-performance trade-off directly addresses a critical constraint in production deployments. This work suggests that future multi-agent systems may not require retraining entire models—instead, efficient meta-skill refinement could enable rapid adaptation. The framework's success across four benchmarks and multiple LLMs provides evidence this isn't a narrow solution but a potentially generalizable paradigm for agent system design.

• →Skill-MAS separates experience retention from model parameter updates, enabling use of powerful frozen LLMs while still improving through learning.
• →The framework achieves significant performance gains across four complex benchmarks while maintaining favorable cost-performance trade-offs.
• →Evolved meta-skills demonstrate strong transferability to unseen tasks and different LLMs, indicating generalizable rather than memorized knowledge.
• →The approach bridges the dilemma between inference-time efficiency and training-time adaptability that constrains current multi-agent system architectures.
• →Multi-Trajectory Rollout combined with Selective Reflection creates an optimization loop that efficiently distills systemic experience into strategic principles.