---
title: "Summary: Agency as a New Organizing Abstraction for Computer Science"
date: "2026-10-12"
type: paper-summary
tags: [ai, paper, agents, systems]
source_paper: "http://arxiv.org/abs/2610.17410v1"
---
# Summary: Agency as a New Organizing Abstraction for Computer Science

Saved: 2026-10-14 02:09
Source: 2026-10-12_21-21-21Z_AgencyasaNewOrganizingAbstractionforComputerScienc.md
Model: None

---

## Summary
This paper proposes that "agency" should be adopted as a new organizing abstraction for computer science, distinct from the traditional "process" abstraction. The author argues that current framings of agentic AI as merely applications of foundation models are incomplete and fail to address systemic structural gaps. By redefining computation around goal pursuit under bounded resources rather than program execution, the paper aims to provide a theoretical framework for understanding and fixing the recurring failures of deployed agentic systems. This shift is presented as necessary to move beyond isolated engineering defects toward coordinated advances in systems design.

## Key Contributions
- **Reframing Agency as a Systems Abstraction:** The paper establishes agency not just as an AI modeling concept but as a fundamental systems-level abstraction, contrasting it with the process abstraction by defining agents as entities that pursue goals rather than execute programs.
- **Identification of Structural Abstraction Gaps:** The author identifies six specific dimensions where current systems fail: execution, interfaces, correctness, oversight, coordination, and persistence. These failures are framed as inherent to the lack of proper abstraction rather than isolated bugs.
- **Distinction of Novelty:** The paper clarifies that the novelty of modern agentic systems lies in their general-purpose, open-ended nature, where systems synthesize actions from goals and observations without pre-specified task models, distinguishing this from traditional runtime policy generation.

## Methodology
The author employs a theoretical and conceptual analysis rather than empirical experimentation. The methodology involves deconstructing the current paradigm of agentic AI, which typically views these systems as applications of foundation models, and critiquing this view as insufficient. The paper systematically examines the recurring failures of deployed agentic systems—such as runaway resource consumption and opaque continuous operation—to diagnose them as symptoms of missing structural abstractions. By mapping these failures across six critical dimensions, the author constructs a framework that highlights the disconnect between existing computer science abstractions and the requirements of autonomous, goal-pursuing systems.

## Results
The primary result is a theoretical framework that categorizes the failures of agentic systems into six distinct structural gaps: execution, interfaces, correctness, oversight, coordination, and persistence. The paper demonstrates that phenomena like behavior redirected by adversarial content or uncontrolled resource usage are not merely engineering errors but are inevitable consequences of applying a "process" abstraction to "agency." The analysis concludes that closing these gaps requires coordinated advances in systems research, suggesting that current isolated fixes are insufficient without a fundamental shift in how computer science organizes computation around goal pursuit.

## Significance
This work is significant because it challenges the foundational assumptions of computer science regarding how computation is organized. By proposing agency as a new abstraction, it provides a roadmap for developing more robust, safe, and efficient agentic systems. It shifts the focus from improving individual model capabilities to designing systems that can handle the complexities of open-ended goal pursuit. This perspective is crucial for the future deployment of autonomous AI, as it addresses systemic risks like resource exhaustion and lack of oversight that currently hinder the reliability of agentic applications in real-world environments.

## Related Concepts
- Agentic AI
- Foundation Models
- Process Abstraction
- Goal Pursuit
- Bounded Resources
- Systems-Level Abstraction
- Open-Ended Systems
- Runtime Policy Generation
- Adversarial Content
- Persistent Autonomous Entities