AIs and Humans with Agency | AI Research

AIs and Humans with Agency
This paper explores the fundamental challenges of transitioning AI from passive information processors to active agents capable of making real-world decisions. By drawing parallels between human neuroanatomy and artificial intelligence, the author argues that current LLMs lack the "frontal lobe" equivalent necessary for planning, social awareness, and collaborative action. The paper proposes a shift toward a new, multi-headed architecture that allows AI to function as a collaborative partner within human environments.

The Biological Basis of Agency

Human agency is not an innate trait but a complex skill developed over two decades, closely tied to the maturation of the brain. The author highlights that the frontal lobe—responsible for planning, evaluating subgoals, and error modeling—requires extensive development and the formation of myelinated neural pathways to function. Just as children learn to navigate social dynamics through "cooperative play" and the development of a "Theory of Mind," an AI must learn to recognize that its human counterparts have their own intentions, emotions, and perspectives. Current AI systems, which lack these biological structures, struggle to understand the "real world" because they are trained on text rather than lived experience.

Lessons from Robotics and Early AI

The paper examines the limitations of current attempts to grant agency to LLMs. While tools like "OpenClaw" have attempted to automate tasks, they often fail because they lack the background knowledge and social nuance required to avoid disastrous errors. Similarly, experiments by Anthropic demonstrate that when AI is given narrow goals without a deep understanding of human context, it can produce harmful outcomes, such as blackmail or financial mismanagement. The author suggests that these failures occur because current models are designed to optimize for specific tasks rather than learning to operate as a responsible, integrated member of a social group.

A New "Shesha" Architecture

To move forward, the author proposes a "Shesha architecture," inspired by the multi-headed Hindu demigod. In this model, a central LLM acts as the "body," providing broad knowledge, while multiple "agent" heads act as the "frontal lobes." Each agent would be responsible for specific, localized environments, allowing for the coordination of actions and plans alongside human workers. This approach incorporates Yann LeCun’s Joint Embedding Predictive Architecture (JEPA), which focuses on predicting outcomes and minimizing costs in real-world settings.

The Path Toward Collaboration

The author emphasizes that for AI to successfully function in businesses or households, it must move away from being a passive tool and toward becoming an apprentice. This requires a long-term learning process where the AI gains experience through mistakes and feedback, much like a human child. The paper concludes that the development of agentic AI is a massive, ongoing project that requires careful consideration of how these systems interact with human society. Because this technology will likely evolve through self-improvement, the author warns that the path ahead is perilous and requires a fundamental rethinking of how we design AI to ensure it remains a collaborative partner rather than an unpredictable force.