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

The concept of artificial agency, particularly in the context of curiosity-driven, resource-bounded AI agents embedded within human-tool systems, aligns with emerging research in agentic AI and embodied intelligence. These agents are designed to go beyond passive response mechanisms by actively perceiving, reasoning, and acting in dynamic environments—both digital and physical—to achieve goals with minimal human intervention .

Curiosity in AI agents is increasingly recognized as a driver for autonomous exploration and learning. While large language models (LLMs) provide a cognitive substrate for goal interpretation and planning, integrating intrinsic motivation mechanisms—such as curiosity—enables agents to seek out novel information, identify knowledge gaps, and initiate actions to improve understanding or performance . This mirrors biological intelligence, where curiosity supports adaptive behavior and long-term learning, especially in open-ended or uncertain environments.

Compression plays a critical role in enabling scalable and efficient agency under resource constraints. Agents must manage memory and computation effectively, particularly when dealing with long interaction histories or complex workflows. Techniques such as retrieval-augmented generation (RAG), KV cache offloading, and hierarchical memory models help compress and organize information across short-term interactions, mid-term planning objectives, and long-term knowledge repositories . External memory systems also allow agents to maintain interaction histories without overwhelming computational resources, supporting personalization and continuity in human-agent collaboration .

Communication is another foundational pillar, encompassing both human-agent and agent-agent interaction. Effective agents must be able to interpret natural language instructions, clarify ambiguities through dialogue, and adapt responses based on user expectations and context . In multi-agent systems, communication enables coordination, task delegation, and collaborative problem-solving—what some describe as an "agentic mesh" where multiple specialized agents work together toward shared outcomes . Furthermore, explainability and transparency in decision-making are essential for trust, safety, and human oversight, particularly when agents operate in high-stakes domains like healthcare or finance .

In embodied AI, these principles converge as agents interact with physical environments using world models—internal simulations that capture object dynamics, spatial relationships, and physical laws . Such models allow agents to predict outcomes, plan actions, and simulate future states in a manner compliant with real-world physics, bridging the gap between abstract reasoning and physical actuation .

Overall, the Artificial Agency Program's focus on curiosity, compression, and communication reflects a shift toward more autonomous, adaptive, and integrated AI systems. These systems are not only capable of executing tasks but also of initiating actions, learning from experience, and collaborating within complex socio-technical ecosystems—a vision increasingly realized through advances in agentic and embodied AI architectures .

The Artificial Agency Program outlines a comprehensive research agenda for developing AI agents that transition from passive processing to embodied, curiosity-driven agency. It posits that true agency emerges from the interplay of three core principles: curiosity, compression, and communication. Rather than relying on static objectives or infinite compute, the program focuses on resource-bounded agents embedded within human-tool systems. These agents operate by actively sensing their environment and utilizing intrinsic motivation (curiosity) to explore and reduce uncertainty, leading to more adaptive and robust behavior in complex, real-world settings.

A key contribution of this work is its theoretical framework linking information compression with agentic behavior. By framing perception and action as a process of compressing environmental data into actionable representations, the authors suggest that agents can achieve high-level sensing and actuation capabilities despite hardware or computational constraints. Furthermore, the program emphasizes the role of communication not merely as language generation, but as a necessary mechanism for coordination between agents and humans. This approach bridges the gap between isolated machine learning models and integrated cyber-physical systems capable of meaningful collaboration.

This research matters significantly as it addresses the current limitations of AI in physical and dynamic environments. By prioritizing resource-boundedness and curiosity, it offers a pathway toward creating AI systems that are not only more efficient but also safer and more aligned with human intent. The vision of agents embedded in human-tool systems promises to enhance human capabilities through improved sensing and automated actuation, marking a critical step toward the realization of practical, embodied artificial intelligence in everyday technology.

`markdown # Summary: Artificial Agency Program – Curiosity, Compression, and Communication in Agents

The Artificial Agency Program (AAP) introduces a framework for curiosity-driven, resource-bounded AI agents designed to operate within human-tool ecosystems. The paper positions these agents as embodied systems that enhance sensing and actuation by leveraging intrinsic motivations (curiosity) to explore environments, compress knowledge efficiently, and communicate insights in a human-aligned manner. Central to the AAP is the idea that agency emerges from the interplay of curiosity—driven by predictive coding and information compression—as a means to optimize resource use, and communication—acting as a bridge between the agent’s internal models and external collaborator needs. The work draws from cognitive science, embodied AI, and information theory to argue that these principles are foundational for scalable, generalizable AI agents.

Key contributions include a formalization of compression-driven curiosity as a mechanism for efficient learning under uncertainty, and a model of communication as a bottleneck that forces agents to distill and transmit only the most relevant information. The paper also proposes a methodology for embedding such agents in mixed human-AI systems, where curiosity guides exploratory behavior, compression ensures robustness under limited resources, and communication enables seamless collaboration. These insights matter because they address critical gaps in current AI systems—particularly the lack of intrinsic motivation, the inefficiency of black-box models in real-world deployment, and the poor integration of AI with human workflows. By grounding agency in curiosity and communication, the AAP offers a path toward more autonomous, interpretable, and collaborative AI agents. `

This summary highlights the technical and conceptual contributions while emphasizing the broader relevance to AI research and real-world applications.