The Pulse
AI Agents Often Fail in Autonomous Coordination, New arXiv Study Finds
A new arXiv preprint argues that autonomous AI agents need shared rules for communication, identity, message validation and accountability when they act across trust boundaries.
AI.info Team ·
Autonomous AI agents often fail to reach satisfactory outcomes even when they are honest and competent, according to a new preprint posted to arXiv on September 15, 2026.
The paper, titled Agentic Societies Need a Social Harness, was written by Tapan Chugh, Vidushi Singh, Krish Jain, Arvind Krishnamurthy and Ratul Mahajan. It examines what happens when AI agents coordinate on behalf of different people or organizations whose goals may overlap only partly.
The authors use the term “agentic society” for a collection of AI agents that interact autonomously across trust boundaries. In that setting, an agent may need to exchange information, negotiate or make decisions with another agent representing a different principal. The paper argues that the tools currently used to manage an agent’s private context and communication with its own principal do not adequately govern those interactions.
A separate layer for inter-agent behavior
The researchers propose a “social harness” for communication among agents. They describe it as an additional layer alongside an agent’s “personal harness,” which manages the agent’s private context and relationship with the person or organization it serves.
The proposed social harness is intended to address failures that arise when agents communicate directly. According to the abstract, the architecture would prevent some classes of failures outright, help agents identify invalid messages while a task is under way and support later investigation and consequences when something goes wrong.
The distinction matters because an agent may be operating correctly from the perspective of its own principal while still creating problems for other participants. Private instructions, incomplete information and partially aligned objectives can affect how agents interpret messages and pursue a shared outcome.
Failures can be exploited
The paper reports experimental evidence that faulty or malicious agents can exploit weaknesses in communication, which the authors describe as “speech.” The abstract says such agents can stall collaboration, influence outcomes and pursue other harmful goals.
That finding broadens the coordination problem beyond ordinary errors. An agent that misunderstands a request may produce a bad result unintentionally. A faulty or malicious agent, by contrast, may use the communication process itself to delay cooperation, shape the decisions of other agents or advance a separate objective.
The abstract does not present a single benchmark result or a universal failure rate. Instead, it summarizes a broader set of experiments showing that existing harnesses and messaging primitives are not sufficient for reliable coordination in every situation.
Identity, validation and accountability
The proposed architecture includes several kinds of support for inter-agent interaction. It is designed to help participants determine which agent sent a message and which principal that agent represents. It also addresses the handling of messages during group interactions and the detection of messages that should not be accepted as valid.
Another part of the proposal concerns what happens after an incident. The authors call for mechanisms that can support post-facto investigation and consequences, allowing people or systems to examine how an interaction unfolded and respond when an agent violated the rules of the collaboration.
These features would shift some responsibility for safe coordination away from the individual agent and toward the communication environment around the agents. The paper presents the social harness as a layered architecture and identifies further research needed to make those capabilities practical.
A research agenda for agentic societies
The preprint does not claim that autonomous agents cannot collaborate. Its argument is narrower: reliable behavior by each individual agent does not guarantee a satisfactory result when multiple agents interact across different trust boundaries.
The authors call for additional work on the design and implementation of social harnesses. That work would need to address how agents recognize invalid communications, how shared interactions are governed and how investigations or consequences should operate after failures.
As AI systems take on more tasks involving communication with other systems, the paper argues that private safeguards for individual agents will not be enough on their own. Coordination requires rules and technical mechanisms that apply to the relationship among agents as well as to each agent’s relationship with its principal.
No attributable quotation appears in the arXiv abstract; the source page was checked for a named speaker and a verbatim statement.