Artificial Agency Program: Curiosity, compression, and communication in agents
Abstract
This paper presents the Artificial Agency Program (AAP), a position and research agenda for building AI systems as reality embedded, resource-bounded agents whose development is driven by curiosity-as-learning-progress under physical and computational constraints. The central thesis is that AI is most useful when treated as part of an extended human--tool system that increases sensing, understanding, and actuation capability while reducing friction at the interface between people, tools, and environments. The agenda unifies predictive compression, intrinsic motivation, empowerment and control, interface quality (unification), and language/self-communication as selective information bottlenecks. We formulate these ideas as a falsifiable program with explicit costs, staged experiments, and a concrete multimodal tokenized testbed in which an agent allocates limited budget among observation, action, and deliberation. The aim is to provide a conceptual and experimental framework that connects intrinsic motivation, information theory, thermodynamics, bounded rationality, and modern reasoning systems
Figures & tables
| Hyp. | Positive evidence | Disconfirmation / boundary case |
|---|---|---|
| H1 | Across tasks, interventions improving learning progress also improve control/empowerment metrics under fixed budgets. | Predictive gains with no corresponding control gain on key tasks; or control gains via exploitation that degrade predictive competence. |
| H2 | Agent spends budget to widen sensors/actions/communication only when long-horizon return improves; rises then saturates. | Agent never invests in interface improvements despite clear gains; or always over-invests regardless of cost. |
| H3 | Under stronger cost/viability constraints, policies shift toward better predictive state and reduced costly reactive plasticity. | Constraint increases merely collapse behavior without improved predictive organization or selective control. |
| H4 | Meta-control beats fixed observe/act/deliberate schedules at equal budget. | No gain over tuned static schedules, implying control overhead dominates. |
| H5 | Private self-communication tokens improve task performance / sample efficienc. | Private channel collapses into redundant verbosity, or latent-only recurrence dominates under equal compute. |