From Agents to Intelligence Fields: A Field-Theoretic Metaphor for the Mode of Existence of Large Language Models

The fluency of large language models (LLMs) invites a persistent metaphor: that an LLM is a person-like agent with beliefs, intentions, and a stable personality. This paper argues that the agent/person metaphor systematically misleads, and proposes an alternative framing: the LLM as an intelligence field—a distributed, high-dimensional space of latent responses that is excited by language rather than a unitary subject that thinks. Building on prior anti-anthropomorphic accounts—role-play (Shanahan et al., 2023), the simulator hypothesis (janus, 2022), and language-models-as-agent-models (Andreas, 2022)—the field view foregrounds three features those accounts underweight: continuity, boundary conditions, and terrain. We argue that prompts function as boundary conditions on the field; that the field is strongly anisotropic, with valleys and walls carved by pretraining and alignment via reinforcement learning from human feedback (RLHF); and that this anisotropy explains why prompt engineering is sometimes powerful and sometimes ineffective. We re-cast prompting, alignment, and multi-turn context in field terms and, crucially, derive a falsifiable operating principle—"survey before you push"—with associated testable predictions, moving the metaphor from perspective to tool. Finally, we distinguish the LLM-as-field from the Agent-as-control-system constructed atop it. The field metaphor is offered as scaffolding, not metaphysics.