The Tao of Agency: Autotelic AI, Embedded Agency an... | AI Research

The Tao of Agency: Autotelic AI, Embedded Agency and Dissolution of the Self explores a shift in artificial intelligence from systems that follow designer-provided instructions to "autotelic" agents—systems that generate, sustain, and revise their own goals. The paper argues that the core challenge of building such agents is not merely the mechanism for goal generation, but the fundamental problem of how an agent defines and maintains the "self" to which those goals are assigned.

Moving Beyond Given Objectives

Most current AI systems rely on a "given-objective" paradigm, where a human designer specifies a reward function or loss function that the agent must optimize. While this has led to significant breakthroughs in fields like game-playing and structural biology, it remains brittle for open-ended tasks. Autotelic AI seeks to move beyond this by allowing the agent to discover its own objectives. The author notes that while intrinsic motivation techniques—such as rewarding curiosity, novelty, or information gain—allow agents to act without external signals, they still rely on a fixed, designer-specified "intrinsic" reward. True autotelic agency requires a more fundamental approach to how an agent chooses what to value.

The Role of Resources and Embodiment

To move toward a more principled way of generating goals, the paper considers how an agent’s physical and computational constraints shape its priorities. An agent is not an abstract machine with infinite resources; it operates within limits of time, energy, memory, and computational power. These constraints act as a "resource-driven prior" that influences which goals are worth pursuing. By grounding goal generation in these physical realities, the agent’s preferences become a reflection of its own embodiment rather than an arbitrary choice by a designer.

The Necessity of the Self

The deepest problem identified in the paper is that for an agent to have a goal, it must first have a boundary that separates "itself" from the "environment." The author uses the concept of a Markov blanket—a statistical boundary that separates internal states from external ones—to explain how an agent individuates itself. However, this individuation is not unique; the same system dynamics can often be partitioned in multiple valid ways, leading to different "candidate selves."

The Paradox of Autotelic Agency

The paper concludes that the act of defining a self is a necessary condition for agency, but it creates a paradox: the agent must believe in its own boundary to act, yet it must also "see through" that boundary to truly understand the world. This leads to a view where the self is not a fixed, metaphysical entity, but a statistical structure that the agent must continuously generate and relativize. The research extends these ideas into a framework that bridges quantum formulations, non-dual philosophical traditions, and practical implementations using large language models.