Work Agents: Systems, Capabilities, and Optimization

Problem

Foundation models are shifting from generating information to executing digital work across tools and applications, but there is no unified view of how agents reliably complete delegated work.

Method

A work-centered, system-level survey. It defines a work agent as a "model + runtime Harness" system that executes tasks in a stateful workspace and produces verifiable results. It characterizes working conditions along two axes, work persistence and work openness, and classifies digital work into five forms by task scope and organization. It then outlines five core capability groups needed for reliable execution and reviews methods for model optimization, Harness optimization and joint optimization of the two. It also analyzes existing evaluation methods and metrics from the perspective of work outcomes.

Results

• Presented a framework that classifies working conditions using two axes, work persistence and openness, and five forms of digital work.
• Outlined five core capability groups needed for reliable work execution.
• Systematized model, Harness and joint optimization, along with outcome-centered evaluation methods.
• Proposed future challenges: continuous work delegation, adaptation to open environments, and the interaction between doing work and improving capabilities.

Limitations

This analysis is based on the abstract. The full text was not available, so the specifics of the five work forms and capability groups, and the scope of the reviewed research, could not be checked. As a survey paper it contains no new experimental results, and the validity of the classification framework has not been empirically tested.

Why it matters

For teams adopting agents for work automation, it is a useful reference for setting design and evaluation criteria from a system perspective that includes not only model performance but also the Harness, the work environment and result verification.

Key terms

Work agent

A system that combines a foundation model with a runtime Harness to carry out delegated work in a stateful workspace and produce verifiable results.

Runtime Harness

An execution framework around the model that manages tool calls, state and execution flow. Even with the same model, results vary depending on the Harness.

Work persistence and openness

Two axes: how long the work continues and must maintain state (persistence), and how open and changing the environment or goals are (openness).

Original

• arXiv abstract
• PDF
• Code
• Project
• Hugging Face

Authors: Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen, Zirong Chen, Zhiyuan Yao, Yifu Guo, et al. · Posted 2026-10-26 · License CC0 1.0 · Based on the abstract

This review is an AI-written summary of the paper. Please treat the original paper as the reference.