Critique of Agent Model

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2606.23991 Date read: 2026-06-24 Connected to: none Escalation: store-only Escalation rationale:

What this is

A definitional critique examining the boundary between automation and agency in LLM systems. The work appears to draw on Cartesian philosophy to clarify what constitutes genuine agency versus marketed "agentic" behavior, addressing both capability engineering and existential risk framing.

What I took from it

The paper identifies a critical semantic slippage in the protocolized systems literature: the term "agent" is applied simultaneously to deterministic automation pipelines and to speculative autonomous systems, collapsing distinct categories. This matters for the new nature research agenda because it suggests that much discourse about "AI agency" may be operating without rigorous grounding in what agency means in computational systems.

The work appears to push back on two opposing camps—both those who market LLM pipelines as agents without justification, and those who project speculative agency onto systems to justify existential concern. If this distinction holds, it implies that many current risk and capability frameworks are debating the properties of objects they haven't adequately defined. This is foundational but not itself a positive account of how agency does emerge in artificial systems, nor a mechanistic law.

Research connections

none yet identified: No active hypotheses or established laws currently in inventory to connect against.

Candidate laws or signals

CL-2606-1: Agency-automation conflation distorts both capability and risk assessment—systems marketed as "agents" may inherit rhetorical autonomy without corresponding decision-theoretic independence, creating false analogies in both engineering and safety discourse.