Offline Skill-Patching Beats DPT-WToM 15.6x on 27% Fewer Tokens

Notes on CoWork-X: Experience-Optimized Co-Evolution for Multi-Agent Collaboration System (arXiv:2602.05004) — Zexin Lin, Jiachen Yu, Haoyang Zhang, Yuzhao Li, Zhonghang Li, Yujiu Yang, Junjie Wang, Xiaoqiang Ji · February 2026

Note published July 24, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

CoWork-X splits agent cognition into two clocks. A Skill-Agent handles every in-episode decision by retrieving from a structured skill library built as hierarchical task networks (HTN) — state queries, operators, and method decompositions compiled into an editable Python file — with zero LLM calls during play. After each episode, a separate Co-Optimizer LLM reads the trajectory logs (runtime failures, 100-timestep stagnation flags, action-type breakdowns) and patches the library under explicit budget constraints and drift regularization, keeping a best-performing historical version on hand for rollback. The setup runs on a two-agent, symmetric version of the DPT-Agent Overcooked-AI-style burger-preparation benchmark, starting from a deliberately broken skill library — syntactically valid HTN rules whose operators skip precondition checks entirely — across 30 episodes.

The split pays off on both axes it targets. CoWork-X's block-mean score climbs from 52.0 (episodes 0–9) to 109.0 (10–19) to 128.0 (20–29), for an overall mean of 96.3, while ReAct (3.2), DPT-WToM (6.2), and Reflexion (−51.0) stay flat or negative over the same span. It does this at 0 online tokens and 2.6 seconds of latency per episode, versus 30,090–79,126 tokens and 67.7–182.3 seconds for the LLM-in-the-loop baselines — about 27x faster than the strongest of them, DPT-WToM. Even counting the offline Co-Optimizer pass, cumulative cost through 30 episodes is 663k tokens for a 96.3 score against DPT-WToM's 903k tokens for a 6.2 score: a 15.6x score gain on 27% fewer tokens. Swapping in seven LLM families as the Co-Optimizer backbone shows the mechanism generalizes — Claude-4.5 and Gemini-2.5-pro hit ceiling scores around 160 within roughly three patch iterations — but not universally: Qwen-3-235B stays stuck near a −60 score, and GPT-5.1/DeepSeek-v3.2 only reach a stable, above-120 result in 2 of 5 runs.