A system-level perspective on digital work agents and their optimization

According to the authors of 0, foundation models are evolving from tools that merely generate information into systems capable of executing digital work across various applications and workflows. The paper provides a unified framework to understand how these agents reliably accomplish delegated tasks, combining foundation models with specific runtime environments to deliver verifiable outcomes 0.

The problem

As stated by the authors 0, existing research lacks a unified perspective on how agents can reliably accomplish delegated work as they transition from generating information to executing actual digital tasks. Without a structured view of working conditions, task scopes, and core capabilities, it remains difficult to systematically optimize and evaluate how these systems perform across stateful workspaces 0.

The idea

To address this gap, the authors introduce a work-centric, system-level perspective 0. They view work agents as systems that combine foundation models with runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes 0. To characterize different working conditions, the framework introduces two complementary dimensions known as work persistence and work openness 0. Furthermore, the paper classifies digital work into five distinct forms based on task scope and organization 0.

The results

Building on this framework, the authors identify five core capability families that are required for reliable work execution 0. The paper examines various approaches to model optimization, harness optimization, and the joint optimization of both components 0. Additionally, the authors analyze existing evaluation methods and metrics specifically from the perspective of work outcomes 0.

Why it matters

According to the authors 0, this paper provides a unified framework for understanding, optimizing, and evaluating work agents. This perspective supports the broader development of systems that are capable of delivering reliable, verifiable, and practically valuable work outcomes 0. It also outlines future research directions, including sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement 0.

Caveats

The source text is an arXiv abstract and does not report experimental performance numbers, benchmark scores, or specific empirical comparisons against alternative systems 0. Readers should note that this is a preprint description of a conceptual framework rather than a peer-reviewed empirical study containing quantitative validation 0.