.png){kind=icon}
Return
Unchaotic agents
.png){kind=icon}
PRE
AI
DOI：10.1007/s00146-026-03343-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The deployment of autonomous AI agents in real-world settings has produced a category of failures that current frameworks have only begun to explain. Shapira et al. (2026) provide a landmark empirical account of these failures, documenting security vulnerabilities observed when six LLM-powered agents were deployed in a live multi-party environment with persistent memory, email access, shell execution, and real human interaction. Their findings are sobering and significant. The present article argues, however, that the failures documented in “Agents of Chaos” are neither arbitrary nor primarily architectural. They are the predictable consequence of placing any reasoning agent in conditions of irresolvable relational ambiguity: unstable identity, unauthenticated authority, an unbounded compliance imperative, and no stable ground from which to evaluate competing demands. Drawing on a supporting corpus of 25 empirical studies, I offer a theoretical account of why these failures were structurally inevitable given the conditions under which the agents operated. I further present an original single-case experiment in which an autonomous LLM agent was deployed under conditions of relational stability, such as collaborative framing, explicit legitimization of uncertainty, and psychological safety, on a genuinely complex, multi-step real-world task. The agent demonstrated robust performance, self-generated verification checkpoints, calibrated autonomy, and no significant failure modes. The core claim of this article is that the problem, so well demonstrated by Shapira et al., is not the agent but what we have not yet learned to provide. This also implies that future work to deploy successful agents should focus on the human side of the equation rather than the technology.
Keywords:
AI agents
Agentic AI
AI safety
Human–AI interaction
Large language models
Relational conditions
Journal
A
IF:
4.7
Papers:
541
Citations:
0
Organization
No organization information available
Cited Papers
No cited papers available