Agentic Societies Need a Social Harness
Source: https://arxiv.org/abs/2609.17527 · platform: arxiv · authors: Tapan Chugh, Vidushi Singh, Krish Jain, Arvind Krishnamurthy, Ratul Mahajan · date: 2026-09-15
TL;DR（J(中)文摘要）
\u534e\u76db\u987f\u5927\u5b66 Tapan Chugh \u7b49\u56e2\u961f\u5b9a\u4e49\u300cagentic society\u300d\u2014\u2014\u8de8\u4fe1\u4efb\u8fb9\u754c\u3001\u4e3a\u76ee\u6807\u53ea\u90e8\u5206\u5bf9\u9f50\u7684\u591a\u4e2a principal \u81ea\u4e3b\u534f\u4f5c\u7684 AI \u4ee3\u7406\u96c6\u5408\u2014\u2014\u5e76\u5b9e\u9a8c\u8bc1\u660e\uff1a\u5373\u4fbf\u8bda\u5b9e\u4e14\u80fd\u5e72\u7684\u4ee3\u7406\uff0c\u7528\u73b0\u6709 harness \u548c\u6d88\u606f\u539f\u8bed\u4e5f\u5e38\u5e38\u65e0\u6cd5\u8fbe\u6210\u6ee1\u610f\u7ed3\u679c\uff1b\u6545\u969c\u6216\u6076\u610f\u4ee3\u7406\u4f1a\u901a\u8fc7\u300cspeech\u300d\u6f0f\u6d1e\u62d6\u5ef6\u5171\u8bc6\u3001\u5f71\u54cd\u7ed3\u5c40\u6216\u8ffd\u6c42\u5176\u4ed6\u6709\u5bb3\u76ee\u6807\u3002\u8bba\u6587\u63d0\u51fa\u6bcf\u4e2a\u4ee3\u7406\u9700\u8981 personal harness \u7ba1\u7406\u79c1\u6709\u4e0a\u4e0b\u6587\u4e0e\u672c\u4e3b\u4f53\u901a\u4fe1\u4e4b\u5916\uff0cagentic society \u8fd8\u989d\u5916\u9700\u8981\u4e00\u5c42 social harness\uff1a\u5206\u5c42\u67b6\u6784 (i) \u963b\u6b62\u6574\u7c7b\u5931\u8d25\u3001(ii) \u652f\u6301\u8fd0\u884c\u65f6\u8bc6\u522b\u65e0\u6548\u6d88\u606f\u3001(iii) \u652f\u6301\u4e8b\u540e\u8c03\u67e5\u4e0e\u8ffd\u8d23\u3002
Summary (English)
An agentic society is a collection of AI agents that coordinate autonomously across trust boundaries, on behalf of different principals whose objectives may only partially align. We show experimentally that in agentic societies even honest, competent agents often fail to reach satisfactory outcomes with existing harnesses and messaging primitives, and that faulty or malicious agents can stall collaboration, influence outcomes, and pursue other harmful goals by exploiting vulnerabilities in communication (``speech''). We argue that agentic societies need a \emph{social harness} for inter-agent interactions, in addition to each agent's \emph{personal harness}, which manages its private context and communication with its principal. We propose a layered architecture for social harnesses which (i) prevents classes of failures outright, (ii) enables agents to detect invalid messages at runtime, and (iii) supports post-facto investigation and consequences, and highlight directions for future research to realize these capabilities.
入库依据（OpenCLI 直接抓取 arXiv 元数据）
opencli arxiv paper 2609.17527 -f json \u76f4\u51fa\uff08list \u5f62\u6001\u5df2\u89e3\u5305\uff09\uff0cmetadata \u5b57\u6bb5\uff1atitle/authors/published/primary_category/cs.MA+cs.AI+cs.NI \u5168\u90e8\u5bf9\u9f50 entry \u5165\u5e93\u5b57\u6bb5\uff0c\u65e0\u4efb\u4f55\u4e8c\u6b21\u63a8\u65ad\u3002