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

2025-04-14 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

SkillMAS is a non-parametric framework designed to improve Large Language Model (LLM) agent systems after deployment by adaptively specializing them. It addresses the common issue of "adaptation decoupling," where skill evolution and multi-agent system (MAS) restructuring are treated as separate problems, leading to bottlenecks and mis-specialization. SkillMAS integrates these two adaptation targets by using Utility Learning to assign credit from verified execution traces, bounded skill evolution to refine reusable procedures without uncontrolled library growth, and evidence-gated MAS restructuring that triggers only when retained failures and Executor Utility indicate a structural mismatch. The framework was evaluated across embodied manipulation (ALFWorld), command-line OS workflows (Lifelong Agent Bench OS Task), and retail workflows (τ-Bench Retail), achieving competitive success rates, such as 94.0% on ALFWorld's unseen split and 76.7% on Lifelong Agent Bench OS Task.

Key takeaway

For AI Architects and AI Engineers designing adaptive LLM agent systems, SkillMAS offers a principled approach to post-deployment specialization. By integrating skill evolution with MAS restructuring through a shared evidence surface, your teams can avoid common pitfalls like context overload and mis-specialization. Consider adopting a similar coupled adaptation strategy to ensure your agent systems improve holistically and efficiently, rather than optimizing skills and organization in isolation.

Key insights

SkillMAS couples skill evolution and MAS restructuring using verified execution traces for adaptive LLM agent specialization.

Principles

• Credit assignment must be from verified execution traces.
• Skill evolution should avoid overwhelming MAS organization.
• Skill evolution and MAS restructuring need shared empirical evidence.

Method

SkillMAS executes episodes, learns Skill and Executor Utility from verified traces, constructs a retained evidence set, then applies bounded skill evolution and evidence-gated MAS restructuring to update the system state.

In practice

• Use Utility Learning to ground skill and executor credit in actual execution.
• Implement bounded skill evolution to prevent uncontrolled skill library growth.
• Trigger MAS restructuring only when evidence indicates structural bottlenecks.

Topics

• SkillMAS Framework
• LLM Multi-Agent Systems
• Skill Co-Evolution
• Post-Deployment Adaptation
• Utility Learning

Best for: Research Scientist, AI Architect, AI Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems — Skill-MAS decouples experience retention from LLM parametric updates using an evolvable Meta-Skill for multi-agent systems.
• SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems — SkillSmith co-evolves agent skills and tools using an ecological utility model and anti-pattern learning to enhance self-improvement.
• Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose — Compositional skill routing for LLM agents requires decomposition, retrieval, and plan composition.
• Self Evolving AI Skills w/ GPT-5.5 (SkillOpt) — SkillOpt enables LLMs to self-evolve skills by optimizing natural language policies through a frozen target model and an external optimizer.
• NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning — NeuroMAS treats multi-agent systems as trainable neural architectures, learning communication and specialization via reinforcement learning.
• Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward — Agent skills modularize LLM capabilities, extending functionality dynamically while introducing significant security challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.