Autotelic AI: The Tao of Agency by Aritra Sarkar, arXiv 2606.19924

Aritra Sarkar submitted The Tao of Agency: Autotelic AI, Embedded Agency and Dissolution of the Self to arXiv as arXiv:2606.19924 on 18 Jun 2026, with version [v1] timestamped Thu, 18 Jun 2026 08:20:33 UTC (31 KB). The paper reframes agents that generate their own goals as a single problem whose deepest difficulty is how the agent forms and relativizes the self to which goals attach.

What does the paper claim?

The paper claims autotelic AI asks agents to discover their own objectives and that the central difficulty is self-generation rather than goal-generation. Sarkar traces consequences of autotelic agency through intrinsic motivation, resource-driven priors, causal-interventional learning, homeostasis, and embeddedness, and finds embeddedness necessary but not sufficient for autotelic agency.

Sarkar argues embeddedness individuates the agent while exposing that the individuation is non-unique, "such that the same dynamics admit many valid partitions, each defining a different candidate self." The abstract frames the problem: an agent must believe in a boundary to act and see through that boundary to understand. The paper therefore shifts the spotlight from how goals appear to how a self is constructed and relativized.

How does the author extend the framework?

Sarkar consolidates the developments into a single framework and extends it along three explicit directions: a quantum formulation where the agent-environment cut becomes physical, a philosophical reading against non-dual contemplative traditions, and a concrete LLM-based agentic instantiation. These three extensions form the paper's forward agenda.

The quantum formulation treats the agent-environment cut as having a physical character, the philosophical reading situates the argument against non-dual contemplative traditions, and the LLM-based instantiation supplies a concrete pathway for implementing agentic behavior within current large language model systems. The abstract lists intrinsic motivation, resource priors, causal-interventional learning and homeostasis as mechanisms the paper explores in service of autotelic emergence.

Why does this matter?

If agents can generate and assign goals only by forming a self, then understanding self-individuation becomes central to any pathway toward robust autotelic AI. The paper reframes safety and design questions: the boundary an agent believes in both enables action and limits understanding, and the non-uniqueness of valid partitions implies multiple competing candidate selves may arise from the same dynamics.

This perspective changes where researchers must focus. Rather than only engineering goal objectives or reward functions, designers may need to consider how embeddedness and mechanisms like homeostasis or causal-interventional learning contribute to which partition of dynamics becomes the operative agent. Sarkar enumerates those mechanisms and locates them within a unified framework, which makes the problem tractable to theoretical and implementational work.

What to watch

Watch for further material following the three extension directions Sarkar names: a formal quantum account of the agent-environment cut, philosophical engagement with non-dual traditions, and details of the promised LLM-based agentic instantiation. The paper is available now on arXiv as arXiv:2606.19924, and an arXiv-issued DOI via DataCite is noted as pending registration.

Additional concrete reference points from the submission: the paper was submitted 18 Jun 2026 and the posted file for [v1] is recorded at 31 KB. Those metadata entries can serve as anchors for readers tracking updates or subsequent versions.

The Tao of Agency reframes autotelic AI around the paradox of agency and self. It does not claim solved implementations, but it maps conceptual terrain and signals three concrete research paths to follow: physics-informed cuts, cross-disciplinary philosophy, and language-model based instantiations.