Xing et al.'s 'Critique of Agent Model' defines genuine AI agency

In a new arXiv paper titled 'Critique of Agent Model,' researchers Eric Xing, Mingkai Deng, and Jinyu Hou tackle one of AI's most pressing conceptual questions: where does automation end and genuine agency begin? Drawing inspiration from Descartes' grounding of agency in independent thought, as well as portrayals of autonomous beings in science fiction, the authors survey the current landscape of LLM-powered 'coding agents,' 'AI co-scientists,' and other agentic tools. They argue that many marketed 'agents' are actually sophisticated workflows orchestrated by external scaffolding—what they call 'agentic' systems—rather than truly autonomous 'agentive' systems whose capabilities arise endogenously.

The paper analyzes agent architectures along five dimensions: goal, identity, decision-making, self-regulation, and learning. For true agency, the authors contend, these structures must be internalized within the system itself. To operationalize this, they propose the Goal-Identity-Configurator (GIC) architecture, which includes hierarchical goal decomposition, identity evolution over time, simulative reasoning grounded in a separately trained world model, learned self-regulation, and self-directed learning from both real and simulated experience. This framework is designed to let systems operate in the open world with true autonomy, not just follow prescribed tasks.

Beyond the technical architecture, the paper offers critical insights on safety and control. The authors discuss auditability, controllability, and safety mechanisms for agentive systems that possess greater autonomy but remain under meaningful human oversight. By clarifying the boundary between agentic and agentive systems, the work helps separate productive automation from speculative existential fears, providing a conceptual toolkit for both building capable systems and understanding what to fear—or not fear—as AI becomes more autonomous.