---
id: 20261110-T0-11
title: "多代理系统该给子代理开多少权限？新方法从Agent Skills中学习精确授权"
title_en: "New Method Learns Precise Subagent Permissions from Agent Skills"
url: https://ai.daily.yangsir.net/daily/20261110-T0-11
issue_date: 2026-11-10
publish_date: 2026-11-09T05:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2611.08905
---

# 多代理系统该给子代理开多少权限？新方法从Agent Skills中学习精确授权

语言代理处理复杂任务时越来越依赖多个子代理分工协作，这带来一个安全问题：不仅要决定每个子代理做什么，还要决定它能访问哪些工具和工作区资源。论文提出从Agent Skills中学习子代理授权的方法，把权限边界从人工规定转为从技能描述中推导。对搭建多代理系统的团队，这意味着权限配置可以随技能定义自动生成，减少越权访问的风险面。

## English Version

**New Method Learns Precise Subagent Permissions from Agent Skills**

As language agents take on complex tasks, they increasingly coordinate specialized subagents—raising a security question not just of what each subagent does but which tools and workspace resources it may access. This paper learns subagent authorization from Agent Skills, deriving permission boundaries from skill descriptions instead of hand-specified rules. For teams building multi-agent systems, permission configuration can follow skill definitions automatically, shrinking the surface for over-privileged access.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2611.08905)

**详情页**：https://ai.daily.yangsir.net/daily/20261110-T0-11

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*