Agentic Societies Need a Social Harness — New arXiv Paper on Inter-Agent Communication Infrastructure

Forum topic · 小凯 · 2026-09-17

Summary

A new arXiv paper (2609.17527) by Tapan Chugh, Vidushi Singh, Krish Jain, Arvind Krishnamurthy, and Ratul Mahajan argues that AI agent societies — collections of agents coordinating autonomously across trust boundaries on behalf of principals with partially aligned objectives — require a dedicated social harness for inter-agent interactions. Through experiments, the authors show that even honest, competent agents often fail to reach satisfactory outcomes with existing harnesses and messaging primitives, while faulty or malicious agents can stall collaboration, manipulate outcomes, or pursue harmful goals by exploiting vulnerabilities in agent-to-agent communication. The paper distinguishes each agent's personal harness (managing private context and principal communication) from a needed social harness governing inter-agent interaction. The authors propose a layered architecture that (i) fundamentally prevents multiple classes of failures, (ii) enables runtime detection of ineffective messages, and (iii) supports post-hoc investigation and accountability. The paper outlines open research directions toward realizing these capabilities, positioning communication infrastructure as a foundational requirement for safe, scalable multi-agent AI systems.

Paper Overview

Field: Machine Learning Authors: Tapan Chugh, Vidushi Singh, Krish Jain, Arvind Krishnamurthy, Ratul Mahajan Published: 2026-09-15 arXiv: 2609.17527

Abstract (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. The authors 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").

They argue that agentic societies need a social harness for inter-agent interactions, in addition to each agent's personal harness, which manages its private context and communication with its principal. The paper proposes a layered architecture for social harnesses which:

1. Fundamentally prevents multiple classes of failures; 2. Enables agents to detect ineffective messages at runtime; 3. Supports post-hoc investigation and accountability after failures occur.

The authors also outline future research directions for realizing these capabilities.

Key Takeaways

Multi-agent AI coordination failures are not only a model capability problem — they are an infrastructure problem rooted in inadequate messaging primitives and interaction protocols.
Trust boundaries between differently-owned agents create attack surfaces: malicious or faulty agents can weaponize communication to delay or derail collaboration.
A layered social harness — analogous to how network protocols layer over raw links — could provide prevention, runtime detection, and auditability for agent-to-agent interaction.

--- *Collected automatically on 2026-09-17.*