# 智能体技能即使与任务完全相关仍可能使 LLM 智能体表现更差

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-17 11:43
- AIHOT 分数：42
- AIHOT 链接：https://aihot.virxact.com/items/cmswp8qhh02pyroztrgrbfbc0
- 原文链接：https://x.com/rohanpaul_ai/status/2089196267930800136

## AI 摘要

一项实证研究发现，与任务完全相关的智能体技能仍可能使 LLM 智能体表现更差。307 例技能引发的失败中，125 例功能性故障有 86 例源于任务实现错误；182 例效率回退中 114 例来自额外流程，过度验证造成 67 例。平台应将加载技能视为改变系统行为，需对照任务要求核查并与无技能运行对比。

## 正文

Very timely paper. An agent skill can be completely relevant to the task and still make the agent worse.

This study compares the same tasks under different skill setups while keeping the model, agent framework, repository, and verifier fixed.

Across 307 confirmed skill-induced failures, the bigger problem was not obviously irrelevant skills. Among 125 functional failures, 86 came from task-implementation faults: the skill pushed the agent to fill a required element incorrectly or omit it entirely.

Cost failures had a similar pattern. Among 182 high-confidence efficiency regressions, 114 came from extra procedure, not just longer prompts. Excessive verification alone caused 67 cases, with skills turning tests, rebuilds, debugging, and checklists into mandatory work.

The practical warning is simple: a skill can be perfectly on-topic and still make an agent worse.

So for agent platforms, loading a skill should be treated like changing system behavior: check it against task requirements, compare it with a no-skill run, and measure the extra actions it induces.

– arxiv. org/abs/2608.11888

Title: "Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents"