TLDR: A new research paper challenges the long-standing ‘agent’ paradigm in AI, arguing that its conceptual ambiguities and human-centric biases may limit the development of advanced intelligent systems. It distinguishes between agentic, agential, and non-agentic systems, critiques LLMs as merely exhibiting ‘algorithmic mimicry’, and proposes a shift towards system-level dynamics, world modeling, and material intelligence. Quantitative analysis reveals a gap between theoretical critique and practical implementation, highlighting opportunities to bridge these areas for more robust and ethical AI.
The concept of an ‘agent’ has been a foundational idea in Artificial Intelligence (AI) for decades, guiding everything from early theories to today’s advanced Large Language Model (LLM) systems. However, a recent paper titled ‘Is the ‘Agent’ Paradigm a Limiting Framework for Next-Generation Intelligent Systems?’ by Jesse Gardner and Vladimir A. Baulin, critically examines whether this agent-centric approach might actually be holding back the development of truly advanced AI.
The authors argue that the persistent ambiguities in defining what an ‘agent’ truly is, coupled with an inherent human-centric bias, could be limiting our understanding and creation of future intelligent systems. To clarify this, they distinguish between three types of systems:
Understanding Different Intelligent Systems
Agentic systems: These are AI systems that are inspired by the idea of agency. They often appear semi-autonomous and goal-directed, like many LLM-based applications. However, they typically lack the deep, intrinsic autonomy found in living organisms.
Agential systems: This category refers to fully autonomous, self-producing, and self-maintaining systems. Currently, biological organisms are the only known examples of truly agential systems, where agency is an intrinsic part of their nature.
Non-agentic systems: These are simply tools or processes designed to perform tasks without giving any impression of agency. They follow direct instructions or fixed algorithms.
The paper suggests that much of what we call ‘Agentic AI’ is a sophisticated imitation. While useful, this framing can obscure the actual computational mechanisms at play, especially in LLMs. The authors highlight challenges in clearly defining and measuring properties like autonomy and goal-directedness, which are central to the agent paradigm.
The Agentic Facade of LLMs and Active Inference
Large Language Models have driven a rapid expansion of ‘Agentic AI’ systems capable of complex tasks. Yet, the paper questions whether LLMs possess genuine agency or merely exhibit sophisticated ‘algorithmic mimicry’ or a ‘simulacra’ of agency. Their impressive performance, it is argued, stems from vast implicit ‘world models’ learned during training, rather than true understanding or intrinsic long-term planning. Attributing human-like qualities like ‘thinking’ or ‘understanding’ to these systems can be misleading and might misdirect research efforts.
Even theoretical frameworks like Active Inference (AIF), which models agents as systems minimizing ‘surprise’ to maintain themselves, are critiqued for potentially introducing anthropocentric biases. While elegant, AIF’s use of human-like cognitive language (e.g., ‘beliefs’, ‘preferences’) might attribute cognitive processes to systems where only complex statistical or physical mechanisms exist. The paper suggests that simpler, non-agentic approaches might be more computationally efficient for achieving adaptive behavior in many contexts.
Shifting Towards System-Level Intelligence
As an alternative, the paper proposes a shift in focus towards frameworks grounded in system-level dynamics, world modeling, and material intelligence. This perspective emphasizes that intelligence, or even agency itself, could be an emergent property resulting from the intricate interplay of numerous components, complex network structures, and continuous environmental interactions, rather than residing solely within discrete ‘agentic’ units.
Key aspects of this alternative framing include:
-
Primacy of World Models and Continuous Interaction: Intelligence is seen as arising from systems that learn, represent, and utilize internal ‘world models’ through ongoing, continuous interaction with their environment. LLMs, for instance, derive their power from implicit world models encoded in their parameters.
-
Materiality and Unconventional Computing: The paper highlights the ‘Matter computes’ hypothesis, suggesting that intelligent behavior can emerge directly from the physical properties and dynamics of the implementing substrate, rather than solely from abstract algorithms. This includes unconventional computing paradigms like reservoir computing or neuromorphic hardware.
Ethical and Practical Considerations
The conceptual ambiguities of the agent paradigm also raise significant ethical and practical concerns. Anthropomorphic projections onto AI ‘agents’ can lead to misplaced trust and ambiguous accountability, creating a ‘moral crumple zone’ where responsibility becomes unclear. The paper discusses dilemmas like Goodhart’s Law (where optimizing a measure can lead to unintended consequences) and Jevons Paradox (increased efficiency leading to increased overall consumption).
A Quantitative Diagnosis of the Field
To support its arguments, the paper includes a quantitative analysis of the AI field’s intellectual structure. This analysis, based on a knowledge graph of concepts from literature, reveals that critiques of the agent paradigm, especially those concerning anthropocentric biases, are now highly influential. It also identifies a significant and persistent gap between high-level theoretical/critical debate and its translation into practical methods and applications.
The authors present an ‘Atlas of Opportunity’, pinpointing promising research frontiers. The strongest opportunities lie in bridging the gap between methods/techniques and critiques/challenges, and between application domains and critiques/challenges. This means developing new methods that directly address the identified problems and rigorously testing existing methods against critical perspectives.
Also Read:
- Navigating the Ethical Landscape of Autonomous AI Agents
- Unpacking AI’s Moral Compass: How Language Models Prioritize Values
Conclusion: Beyond the Agent Metaphor
The paper concludes that while agentic models have heuristic utility, an over-reliance on this single metaphor may constrain scientific imagination. Progress towards truly general intelligence might require discovering principles fundamentally different from those governing individual human agency. The future of AI, it suggests, may lie in systems where intelligence emerges organically from interaction, self-organization, and material properties, moving beyond engineering AI in our own cognitive image. For a deeper dive into this critical re-evaluation, you can read the full research paper here.
