Published July 29, 2026 | Version v1

Agents Do Not Exist: Agency as a Derived Abstraction over Human–Model–Tool Interaction

  • 1. RAI - Radiotelevisione Italiana

Description

The language of “agents” has become pervasive in descriptions of large-language-model systems. A model equipped with tools, memory, iterative execution, and access to an environment is routinely described as an agent; systems containing several such loops are correspondingly described as multi-agent systems. This paper argues that the agent is not a necessary primitive in the computational ontology of such systems. We introduce a minimal ontology containing humans, language models, tools, memories,
environments, messages, and states, together with relations describing communication and state transition. The human user is explicitly included within the ontology rather than treated as an external source of tasks. We show that behaviors ordinarily attributed to an agent—including planning, tool use, observation, memory, reflection, delegation, and iterative problem solving— can be represented as compositions of message-producing and state-transforming interactions among these components.
On this basis, we formulate an agent-elimination principle: under an observational semantics, an architecture whose specification contains an agent primitive can be transformed into an interaction architecture without such a primitive while preserving its externally observable traces. The claim “agents do not exist” should therefore be understood as an eliminativist claim about computational ontology, not as the assertion that the word “agent” is useless. Agency remains useful as a higher-level description of certain interaction patterns, much as high-level programming constructs remain useful despite their elimination during compilation. This reframing has consequences for AI architecture, evaluation, alignment, human–AI interaction, and multi-agent systems. In particular, once the human is placed inside the interaction ontology, apparently autonomous agency can often be understood as a property of a larger human–model–tool system rather than of a privileged computational individual.

Files

AgentsDoNotExist.pdf

Files (250.8 kB)

Name Size Download all
md5:5da9053070a9c596757c98e24ef20d97
250.8 kB Preview Download

Additional details

References

  • Rodney A. Brooks. Intelligence without representation. Artificial Intelligence, 47(1–3):139–159, 1991
  • Daniel C. Dennett. The Intentional Stance. MIT Press, Cambridge, MA, 1987
  • Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Pearson, 4 edition, 2021
  • Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023.