Reorganize a team without touching its skills, and you lose 26 points to doing nothing at all

Notes on SkillMAS: Skill Co-Evolution with LLM-based Multi-Agent System (arXiv:2605.09341) — Shuaidong Pan, Yixian Liu, Jiaye Gao, Te Gao, Weiwen Liu, Jianghao Lin, Zhihui Fu, Jun Wang, Weinan Zhang, Yong Yu · May 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SkillMAS couples two adaptation loops that agent frameworks usually run separately: skill evolution (refining a library of reusable procedures) and MAS restructuring (changing which executor handles what). Both loops draw on the same evidence surface — verified execution traces — rather than retrieval frequency. Utility Learning credits only skills and executors that were actually exercised in a verified trace, not ones merely pulled into context, and tracks a separate Executor Utility signal per team member. Skill edits stay bounded: clean successes turn into reusable motifs, failures enter a patch pool only when analysis isolates one dominant editable cause, and proposals get deduplicated and pruned instead of piling up. Restructuring is evidence-gated too — it fires only when retained failures and Executor Utility point to a genuine organizational mismatch, capped at one edit per round. Testing spans ALFWorld (embodied manipulation), the OS task from Lifelong Agent Bench (command-line execution), and τ-Bench (retail workflows).

Coupling beats running either loop alone, and the ablation tells you more than the headline score. On ALFWorld's 134-task unseen split, full SkillMAS reaches 94.0% versus 93.0% for the strongest single-loop baseline (Traj-Bootstrap). Strip out skill evolution and restructure the team anyway: success falls to 50.0%, worse than the 76.1% a seed system gets by adapting nothing at all. Evolve only the skills and leave the team structure alone, and it climbs to 68.7% — better, but still short of full coupling. On the Lifelong Agent Bench OS task SkillMAS hits 76.7% against a 70.0% best baseline; on τ-Bench retail it holds a single executor and lifts success from 43/74 to 51/74 through skill adaptation alone, while a forced-expansion probe that adds a second executor drops training success to 32/74 — the evidence gate correctly declining an unjustified reorganization.