lessw-blog introduces a foundational project to categorize how different scientific fields define agency, treating it as a frame-dependent compression strategy to inform future AI development.
\nThe Hook
In a recent post, lessw-blog discusses the initiation of a comprehensive project aimed at building a taxonomy of agents across multiple scientific disciplines. This publication serves as both an introduction to an ambitious theoretical endeavor and a direct request for community feedback on how to structure its future outputs.
The Context
The concept of agency is central to the development of modern artificial intelligence. However, as AI systems become increasingly sophisticated and autonomous, the industry faces a significant semantic and conceptual hurdle: the term agent is used ubiquitously but often inconsistently. Historically, fields ranging from control theory and microeconomics to evolutionary biology and cognitive science have all evolved distinct, specialized conceptions of what constitutes an agent. Understanding these diverse definitions is critical right now. AI practitioners need robust, well-defined frameworks to design, evaluate, and interact with autonomous systems safely and effectively. When interdisciplinary insights remain siloed behind field-specific jargon, the AI community misses out on decades of prior research into goal-directed behavior, optimization, and system dynamics.
The Gist
To address this fragmentation, lessw-blog explores the idea that agency is not an absolute, objective physical property of a system. Instead, it is best understood as a frame-dependent compression strategy utilized by observers. This perspective suggests that depending on the specific prediction challenges a scientific field faces, observers develop unique ways to compress complex, underlying behaviors into the simplified concept of an agent. For example, an economist's rational utility maximizer is a different compression than a biologist's fitness-maximizing organism or a control theorist's feedback loop. By mapping these different compressions across disciplines, the author aims to chart a phylogeny of agents. This mapping will help clarify why certain fields model agents the way they do and how those models succeed or fail under different conditions. Ultimately, this foundational work is designed to bridge conceptual gaps, providing AI researchers with a highly nuanced, context-aware understanding of agent design. The current post outlines this theoretical approach and invites readers to contribute their thoughts on what specific deliverables would be most valuable to the community.
Conclusion
For researchers, developers, and theorists interested in the foundational definitions of autonomous systems, this project offers a highly compelling lens through which to view artificial intelligence. Understanding the historical and cross-disciplinary context of agency is a vital step toward building more capable and aligned AI systems. We highly recommend engaging with this foundational work. Read the full post to explore the proposed taxonomy, understand the compression strategy framework, and contribute your own feedback to the ongoing discussion.
\n\n