SkillMAS: Skill Co-Evolution with LLM-based Multi-Agent System

2026-05-10

AI Summary

This work addresses the challenge that existing large language model–based multi-agent systems struggle to jointly optimize skill evolution and system architecture after deployment, leading to organizational bottlenecks, context overload, and misaligned specialization. To overcome this, we propose SkillMAS, a framework that uniquely integrates both optimization dimensions: it evaluates skill utility through execution-trajectory–driven learning, constrains skill library growth via bounded evolution, and dynamically triggers evidence-gated structural reconfiguration based on utility and failure signals. Built upon a nonparametric multi-agent architecture, SkillMAS enables verifiable, adaptive specialization updates. Empirical results across embodied manipulation, command-line execution, and retail workflow tasks demonstrate significant performance gains, while clearly elucidating the mechanisms underlying post-deployment attribution, updating, and application of specialized capabilities.

Abstract

Large language model (LLM) agent systems are increasingly expected to improve after deployment, but existing work often decouples two adaptation targets: skill evolution and multi-agent system (MAS) restructuring. This separation can create organization bottlenecks, context pressure, and mis-specialization. We present SkillMAS, a non-parametric framework for adaptive specialization in multi-agent systems that couples skill evolution with MAS restructuring. SkillMAS uses Utility Learning to assign credit from verified execution traces, bounded skill evolution to refine reusable procedures without unfiltered library growth, and evidence-gated MAS restructuring when retained failures and Executor Utility indicate a structural mismatch. Across embodied manipulation, command-line execution, and retail workflows, SkillMAS is competitive under the reported harnesses while clarifying how post-deployment specialization is attributed, updated, and applied.

Problem

• skill evolution
• multi-agent system
• adaptation
• specialization
• LLM agent

Innovation

• Skill Evolution
• Multi-Agent System Restructuring
• Utility Learning
• Non-parametric Framework
• Evidence-Gated Adaptation