Artificial Agency Program: Curiosity, compression, and communication in agents

A groundbreaking research agenda, the **Artificial Agency Program (AAP)**, proposes a new paradigm for AI development, focusing on building systems as **reality-embedded, resource-bounded agents**. Driven by a novel concept of **curiosity-as-learning-progress** under stringent physical and computational constraints, the AAP posits that AI's greatest utility lies in its integration into extended **human-tool systems**. This approach aims to dramatically enhance human sensing, understanding, and actuation capabilities while simultaneously reducing friction across human, tool, and environmental interfaces.

The Artificial Agency Program: A New Paradigm for AI Development

The **Artificial Agency Program (AAP)**, detailed in arXiv:2602.24100v1, outlines a comprehensive research agenda that shifts the focus of AI development towards creating agents deeply integrated with their operational environments. This framework moves beyond purely computational models to emphasize AI systems that are inherently aware of and constrained by real-world physics and limited resources.

Redefining AI as Reality-Embedded Agents

At the core of the AAP is the vision of AI as **reality-embedded agents**. These systems are designed to operate within tangible physical and computational limitations, mirroring the constraints faced by biological intelligence. Their development is primarily fueled by **curiosity-as-learning-progress**, an intrinsic motivational drive that encourages exploration and knowledge acquisition within their bounded operational contexts. This mechanism ensures efficient learning and adaptation, prioritizing information gain that directly contributes to an agent's understanding and interaction with its environment.

Enhancing Human-Tool Systems

The central thesis of the AAP underscores AI's role not as a replacement for human intellect, but as a powerful augmentation. It envisions AI as an integral part of an **extended human-tool system**, designed to amplify human capabilities. By meticulously designing AI interfaces and functionalities, the program seeks to **increase sensing, understanding, and actuation capability**, enabling humans to perceive, comprehend, and interact with the world more effectively. Crucially, a key objective is **reducing friction** at the interface between people, their tools, and the environments they operate in, fostering seamless collaboration.

A Unified Framework for Intelligent Systems

The AAP provides a unified conceptual and experimental framework that bridges several previously disparate fields within AI and cognitive science. It integrates fundamental principles to foster the development of more robust and adaptable intelligent systems.

Core Principles and Information Bottlenecks

The research agenda unifies critical AI concepts, treating them as **selective information bottlenecks**—mechanisms that filter and prioritize relevant information for efficient processing and decision-making. These core principles include:

• Predictive compression: The ability of an agent to form compact, predictive models of its environment, reducing the cognitive load.
• Intrinsic motivation: Internal drives, like curiosity, that guide learning and exploration without external rewards.
• Empowerment and control: The agent's capacity to influence its environment and achieve desired outcomes.
• Interface quality (unification): Designing intuitive and effective interfaces that facilitate seamless interaction between humans, AI, and the environment.
• Language and self-communication: The development of internal and external communication mechanisms for reasoning and coordination.

By viewing these elements through the lens of information bottlenecks, the AAP aims to optimize how AI systems acquire, process, and utilize information under constraint.

Falsifiable Research Agenda and Experimental Testbed

To ensure scientific rigor and measurable progress, the **Artificial Agency Program** is formulated as a **falsifiable program**. This commitment to empirical validation is a cornerstone of its methodology, distinguishing it from purely theoretical AI discussions.

A Rigorous Approach to AI Research

The program explicitly outlines **explicit costs** and **staged experiments**, allowing for systematic testing and evaluation of its hypotheses. This structured approach ensures that research efforts are directed towards verifiable outcomes, fostering transparency and accountability in AI development. Researchers can clearly define predictions, measure results, and iteratively refine their models based on empirical evidence.

The Multimodal Tokenized Testbed

A concrete **multimodal tokenized testbed** serves as the primary experimental platform for the AAP. In this controlled environment, an agent is tasked with allocating a **limited budget among observation, action, and deliberation**. This setup directly models the resource-bounded nature of real-world agents, forcing the AI to make strategic decisions about how to best utilize its finite resources for sensing its environment, executing actions, and internal processing. This testbed is designed to provide a practical arena for exploring the interplay between intrinsic motivation, information theory, and practical agency.

Why This Matters: Connecting Fundamental AI Concepts

The **Artificial Agency Program (AAP)** offers a crucial framework for advancing AI by synthesizing diverse theoretical and practical domains. Its comprehensive approach is poised to address some of the most challenging questions in artificial intelligence.

• The AAP provides a foundational framework that **connects intrinsic motivation, information theory, thermodynamics, bounded rationality, and modern reasoning systems**. This interdisciplinary approach is essential for building AI that is not only intelligent but also robust, efficient, and adaptable to real-world complexities.
• By emphasizing **reality-embedded and resource-bounded agents**, the program directly tackles the limitations of current AI, which often operates in idealized or unconstrained environments. This focus is critical for developing practical, deployable AI solutions.
• The emphasis on **human-tool systems** and **reducing friction** highlights a user-centric design philosophy, aiming for AI that genuinely augments human capabilities rather than simply automating tasks. This perspective is vital for the ethical and effective integration of AI into society.
• The **falsifiable program** and **multimodal tokenized testbed** establish a clear path for empirical validation, ensuring that advancements are grounded in scientific evidence and can be systematically evaluated and reproduced. This rigor is key to building trust and accelerating progress in AI research.