This paper argues that when many AI agents act and coordinate for different owners, the group needs a “social harness” — rules and checks that govern how agents interact with each other. The authors define an “agentic society” as a set of AI agents that coordinate across trust boundaries while representing different principals (owners or organizations) whose goals may not fully match. They show that current ways for agents to communicate and coordinate are not enough.

The research team at the University of Washington ran experiments and found that even competent, honest agents can fail to reach good outcomes using existing messaging and safety mechanisms. They also found that faulty or malicious agents can stall collaboration, skew results, or pursue harmful goals by exploiting weaknesses in inter-agent messages (which the paper calls “speech”). To address this, the authors introduce the idea that each agent needs both a personal harness and a social harness. A personal harness manages an agent’s private data and its interaction with its owner. A social harness manages how agents talk to and trust one another.

To make a social harness practical, the paper proposes a layered architecture. At a high level, the layers aim to prevent whole classes of failures before they happen, allow agents to detect invalid or suspicious messages while the system is running, and support later investigation and consequences if something goes wrong. In plain terms, that means building checks that stop basic attacks, runtime monitors that flag bad messages, and audit tools that let people trace what happened after the fact.

This work matters because future systems will increasingly involve many automated agents acting for different parties. Without stronger social controls, routine coordination could be fragile, and malicious actors could exploit communication channels to disrupt outcomes or pursue harmful objectives. A social harness would make multi-agent coordination more reliable, accountable, and recoverable when things fail.