Critique of Agent Model: Distinguishing Agentic Tools from Agentive Autonomy

Forum topic · 小凯 · 2026-06-27

Summary

A June 2026 paper by Eric Xing, Mingkai Deng, and Jinyu Hou (CMU, MBZUAI, Petuum) titled "Critique of Agent Model" (arXiv:2606.23991) draws a sharp line between two paradigms. Today's "Agentic" systems—Claude Code, Cursor, AutoGPT—derive their capability from engineered workflows, system prompts, and tool protocols rather than from endogenous intelligence, making them sophisticated puppets. The authors coin "Agentive" to describe true agents whose goals, identity, decision-making, regulation, and learning are internally generated and self-evolving. They formalize this across five dimensions: hierarchical goal decomposition, adaptive self-model (with a Fast-Slow Learning theorem), world-model-based counterfactual planning, a learned Configurator (System III) for meta-decisions, and self-directed training with the world model f separated from the policy π. The proposed GIC architecture integrates Belief Encoder, Goal Decomposer, Identity Evolver, Configurator, Simulative Planner, and Actor. The paper also reframes safety around auditability, controllability, and oversight, and raises philosophical questions about whether machine agency differs fundamentally from biological agency.

Key points

Two paradigms defined: *Agentic* systems (today's LLMs wrapped in scaffolding, prompts, MCP, plan-then-act loops) vs. *Agentive* systems whose goals, identity, decisions, regulation, and learning are endogenous.
Five-dimension critique of current agent design:
*Goal*: external short-term instructions → internal long-term objectives with a hierarchical decomposer δ.
*Identity*: static system prompts → continuously evolving self-model via fast test-time adaptation plus slow retraining (Theorem 1: Fast-Slow Learning Dominates Slow-Only Learning).
*Decision-Making*: black-box policy producing CoT-like tokens → counterfactual planning grounded in an explicit world model f (Theorem 2: any base policy can be augmented to at least match itself using a reasonably accurate world model).
*Self-Regulation*: fixed workflow → learned Configurator κ (System III) that decides whether to act, continue, re-plan, revise goals, or even return to simulation for practice.
*Learning*: human-scheduled training → self-directed learning where the world model f is trained for prediction fidelity and the policy π is trained for goal achievement, kept separate to avoid conflating "what I want" with "what happens."
GIC architecture blueprint: Belief Encoder (h), Goal Decomposer (δ), Identity Evolver (ι), Configurator (κ), Simulative Planner (π_f), Actor (α). Training proceeds Ground School → Simulator RL → Real-World Deployment, mirroring pilot training.
Implication for the field: progress driven by stronger base models and richer scaffolding remains on the Agentic axis and will not automatically reach Agentive autonomy; scaffolding, however elaborate, still leaves the strings in human hands.
Safety reframed: Auditability, Controllability (with hard top-level constraints), and Safety-through-Oversight rather than through tighter external scripting.
Philosophical hook: agency is treated as a structural property (whether organizational capacity is internal or external), not a metaphysical one—even a system originally created by humans can exhibit genuine agency if all five dimensions are endogenous. Citing Descartes (*Cogito, ergo sum*) and *Blade Runner*'s Replicants.

Reference

Xing, E., Deng, M., Hou, J. (2026). *Critique of Agent Model.* arXiv:2606.23991.
Prior: Xing et al. (2025). *Critiques of World Models.*
Related: Deng et al. (2026). *Efficient Agentic Reasoning Through Self-Regulated Simulative Planning.*

Critique of Agent Model: Distinguishing Agentic Tools from Agentive Autonomy

Summary

Key points

Reference

Tags