Agentic Societies Need a Social Harness
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. Experiments show 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. The paper argues 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. A layered architecture for social harnesses is proposed which prevents classes of failures outright, enables agents to detect invalid messages at runtime, and supports post-facto investigation and consequences.
The paper introduces the concept of an agentic society and demonstrates through experiments that current agent harnesses and messaging primitives are insufficient for reliable coordination among autonomous agents with partially aligned objectives. It proposes a social harness as a necessary additional layer, distinct from the personal harness, to manage inter-agent interactions. The proposed layered architecture aims to prevent failures, detect invalid messages at runtime, and enable post-facto investigation and consequences.
The proposed social harness architecture consists of three layers: prevention of failure classes, runtime detection of invalid messages, and post-facto investigation with consequences. This suggests a shift from purely individual agent safety to systemic safety mechanisms that govern communication protocols and enforce norms across agents with different principals.
As multi-agent systems become more prevalent in enterprise and consumer applications, the lack of robust inter-agent communication standards could lead to coordination failures and security vulnerabilities. This research highlights an emerging need for infrastructure that provides social governance for agent interactions, potentially creating a new category of middleware or protocol layers.
Organizations deploying multiple AI agents from different vendors or with different objectives may require social harness solutions to ensure reliable and secure coordination. This could lead to new products or services focused on agent communication governance, auditing, and failure prevention.
Future research directions include realizing the capabilities of the social harness, such as formal verification of communication protocols, reputation systems, and accountability mechanisms. Observable next signals would be publications or products implementing social harness components, or standards bodies addressing inter-agent communication.