The Tao of Agency: Autotelic AI, Embedded Agency and Dissolution of the Self

Researchers explore autotelic AI systems that generate their own goals rather than pursuing designer-specified objectives, introducing a framework that examines how agents define their boundaries and selfhood. The work reveals that agent individuation is non-unique—multiple valid partitions of agent-environment dynamics exist—creating a fundamental paradox: agents must believe in their own boundaries to act while transcending those boundaries to understand. The framework extends into quantum formulations and contemplative philosophy, with practical LLM-based implementations.

This arXiv paper addresses a foundational challenge in AI development that moves beyond traditional goal-specification architectures. Rather than treating objectives as external inputs, autotelic systems must develop intrinsic motivation and self-directed learning mechanisms. The research synthesizes multiple disciplines including reinforcement learning theory, systems philosophy, and quantum mechanics to examine how autonomous agents construct identity boundaries.

The core insight—that agent individuation lacks uniqueness—represents a significant departure from classical assumptions about machine agency. Multiple valid partitions of the same agent-environment system create ambiguity about what constitutes the "self" pursuing goals. This connects to broader trends in AI safety and interpretability, where understanding agent boundaries becomes critical for alignment and control. The paper acknowledges that embeddedness (agent integration within an environment) is necessary but insufficient for true autotelic behavior.

For the AI development community, this framework has practical implications for building more adaptive and self-improving systems. The LLM-based instantiations suggest researchers are moving from theoretical models toward testable implementations. However, the non-uniqueness problem creates potential challenges for verification and safety—if multiple valid self-definitions exist for the same system, ensuring predictable, aligned behavior becomes substantially more complex.

The philosophical grounding in non-dual traditions suggests the researchers anticipate that consciousness and agency problems may require paradigmatic shifts beyond classical computational models. Future work likely involves empirical validation of these theoretical predictions through increasingly sophisticated agent systems, with particular attention to how boundary conditions affect behavior reliability and interpretability.

• →Autotelic AI agents must generate both their own goals and their own self-definitions, creating a fundamental non-uniqueness problem in agent individuation.
• →The non-dual philosophical framework suggests consciousness and agency may require transcending traditional subject-object boundaries in AI architecture.
• →Agent embeddedness within environments enables autonomy but fails to uniquely specify agent boundaries across multiple valid system partitions.
• →Practical implementations using large language models are testing whether theoretical frameworks for self-directed agency translate to functional systems.
• →The framework's quantum formulation points toward potential connections between physical measurement problems and artificial agent boundary definition.